BEGIN:VCALENDAR VERSION:2.0 PRODID:-//HasGeek//NONSGML Funnel//EN DESCRIPTION:Built for humans. Now rebuilding for agents. X-WR-CALDESC:Built for humans. Now rebuilding for agents. NAME:The Fifth Elephant 2026 Annual Conference X-WR-CALNAME:The Fifth Elephant 2026 Annual Conference REFRESH-INTERVAL;VALUE=DURATION:PT12H SUMMARY:The Fifth Elephant 2026 Annual Conference TIMEZONE-ID:Asia/Kolkata X-PUBLISHED-TTL:PT12H X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT SUMMARY:Check-in DTSTART:20260731T031500Z DTEND:20260731T034500Z DTSTAMP:20260728T181111Z UID:session/M1cRnPhZG3PygYYC3Lvvzt@hasgeek.com SEQUENCE:3 CREATED:20260722T162709Z LAST-MODIFIED:20260722T162722Z LOCATION:Bangalore & Hybrid ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Check-in in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Introduction to The Fifth Elephant 2026 Annual Conference DTSTART:20260731T034500Z DTEND:20260731T035500Z DTSTAMP:20260728T181111Z UID:session/MKr5fi2mquENK4bAoDXjeJ@hasgeek.com SEQUENCE:2 CREATED:20260722T162756Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260722T162928Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Introduction to The Fifth Elephant 2026 Annual Conference in A uditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Optimized AI inference with llm-d and a case-study on sovereign AI on heterogenous GPU cluster DTSTART:20260731T040000Z DTEND:20260731T043000Z DTSTAMP:20260728T181111Z UID:session/B25DdZHvTZu7NshmS6KR7g@hasgeek.com SEQUENCE:16 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T101704Z DESCRIPTION:# Introducing llm-d\nllm-d (llm-d.ai) is a high-performance di stributed inference serving stack optimized for production deployments on Kubernetes It provides a transparent routing layer that sits between the c lient and vLLM serving pods\, making key scheduling decisions at request g ranularity:\n- **Prefix-cache-aware routing** — routes requests to pods whose KV-cache already holds shared prefix content (e.g.\, system prompts\ , few-shot examples)\, avoiding recomputation and reducing TTFT.\n- **Load -aware dispatch** — balances requests across pods based on real-time que ue depth and concurrency\, not round-robin or random assignment.\n- **Opti mized Networking** — integrates seamlessly with NIXL\, UCCL and DeepEp t o provide efficient KV-Cache transfer and Expert Parallelism.\n- **Vendor- agnostic scheduling** — works identically across NVIDIA\, AMD\, and Inte l accelerators\; no per-vendor configuration changes.\n- **Drop-in deploym ent** — deploys as a sidecar or gateway atop any existing vLLM deploymen t with zero application-level code changes.\n\n\n## Case study: NxtGen sov ereign cloud\nMost production inference clusters today are single-vendor — not because it is optimal\, but because it is the simplest way to set things up. Real fleets are accumulating heterogeneity through procurement cycles\, supply constraints\, and the widening cost gap between accelerato rs. The open question is whether a single Kubernetes-native serving layer can take a heterogeneous GPU fleet and beat plain k8s round-robin on throu ghput and time-to-first-token\, with no application-level changes.\n\nWe b enchmarked llm-d on the NxtGen sovereign cloud's 3-vendor cluster: 4× NVI DIA H100-NVL + 8× AMD MI325X + 8× Intel Gaudi3 over a shared 100 G RoCE fabric\, serving ibm-granite/granite-4.1-8b and sarvamai/sarvam-30b. Acros s single-vendor pools (NVIDIA-only\, AMD-only\, Gaudi-only) and heterogene ous pools (NVIDIA+AMD\, NVIDIA+AMD+Gaudi)\, llm-d's prefix-cache-aware rou ting delivers +25 to +91% throughput and 5–22× better TTFT vs plain Kub ernetes round-robin — and the advantage grows with pool size and heterog eneity. The biggest win is on the 20-pod 3-vendor pool\, where llm-d hits +91% throughput at the same load.\n\nResources : https://llm-d.ai/blog/het erogeneous-inference-3-vendor-sovereign-cluster\n# Take-aways\nHeterogeneo us GPU fleets stop being a tax once routing is cache- and load-aware. A si ngle Kubernetes serving layer can absorb NVIDIA + AMD + Intel concurrently and beat round-robin by close to 2× on throughput\, with the same pods\, same vLLM\, same flags — only the routing layer differs. The win is una mbiguously attributable to llm-d's prefix-cache-aware router. Sovereign an d on-prem inference is now operationally viable on mixed-vendor hardware. Procurement no longer has to align with a single vendor's roadmap to get g ood aggregate throughput\; older accelerators can absorb low-priority work loads while premium hardware handles latency-sensitive paths.\n# Audience\ nPlatform and SRE teams running on-prem / sovereign / hybrid-cloud LLM inf erence\; ML infrastructure engineers evaluating Kubernetes-native serving stacks (vLLM\, llm-d\, KServe)\; teams considering or already running hete rogeneous GPU fleets (NVIDIA + AMD + Intel) and worried about how to sched ule across them\; sovereign-cloud and regulated-industry teams (BFSI\, gov ernment\, healthcare) who need on-prem inference and cannot rely on hypers caler-only stacks\; open-source contributors interested in the llm-d / vLL M / Gateway-API Inference Extension projects.\n\n# Draft Slides\nPlease ac cess the draft slides [here](https://drive.google.com/file/d/1F3FCKHJubOKT M3g-5gg04YNAMRfqB9LI/view?usp=share_link)\n\n\n# Bio\nPravein Govindan Kan nan is a Staff Research Scientist at IBM Research working on Systems and N etworking for AI Inference. He contributes to open-source projects like ll m-d\, UCCL and NIXL.\nLinkedIn : https://www.linkedin.com/in/praveingk/\nX /Twitter: https://x.com/praveingk\n\n\n\n\nJayanth Babu Reddy is a Princi pal Architect in NxtGen Cloud Technologies. \n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T133259Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/optimized-ai-inference -with-llm-d-and-a-case-study-on-sovereign-ai-on-heterogenous-gpu-cluster-B 25DdZHvTZu7NshmS6KR7g BEGIN:VALARM ACTION:display DESCRIPTION:Optimized AI inference with llm-d and a case-study on sovereig n AI on heterogenous GPU cluster in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Breaking the trilemma: serverless data platform in your own cloud account DTSTART:20260731T040000Z DTEND:20260731T043000Z DTSTAMP:20260728T181111Z UID:session/RoeFyYpY3jR1Uvw8QfUMoh@hasgeek.com SEQUENCE:17 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T101818Z DESCRIPTION:## Description\n\nIf you run data workloads\, you've probably hit the same wall we did. You want compute that stays in your own cloud ac count\, runs fast\, and stays cheap — and sooner or later\, someone tell s you to pick two. This is a hard problem we solved for our own platform\, and this talk shares those learnings around the engineering it takes to g et all three.\n\nWe'll treat it as a trilemma between ease-of-use\, perfor mance\, and cost\, and use serverless Apache Spark on Kubernetes to illust rate. First\, the economics: across different Spark engines\, we compare s elf-managed deployment against using vendor-managed serverless deployment and show how running compute yourself comes out cheaper\, and we'll walk t hrough exactly why. But\, that cost efficiency and speed gained can come a t the expense of complexities of self-managing workloads.\n\nThen we'll go deep on the infrastructure challenges we had to build and solve to make a genuinely easy serverless experience run inside your own cloud account — remote upgrades\, hardened multi-tenancy\, reliable metering\, and the numerous specific problems on Kubernetes infrastructure that no vendor we looked at had solved. For each one\, we'll cover the design choices and t he tradeoffs\, so you can take the same patterns back to your own platform .\n\nI'll close on why this matters more now. As AI makes software integra tions and even maintenance cheap\, the value of labelling things "fully ma naged" is eroding fast. The lasting differentiator is the hard part it can 't fake: running the workload efficiently with less toil. That's also wher e we're headed next — agents that operate the data platform itself.\n\n# # Takeaways\n\n- What the "serverless tax" really costs you — not just t he cloud discount the vendor captures\, but the control you give up over w here your data runs and how the workload is tuned — and how to weigh it for your own workloads.\n- The economic case for keeping compute in your o wn account\, built so it holds even if you assume every engine performs id entically.\n- The specific infrastructure problems that make in-VPC server less hard\, and how they can be solved.\n- A way to weigh Kubernetes deplo yment models for Spark against the three axes that actually conflict: ease -of-use\, performance\, and cost.\n\n## Who should attend?\n\nPlatform and data engineers building or buying managed compute\, and the people who si gn off on those invoices.\n\nThe economics need no deep Kubernetes backgro und to follow\; the infrastructure section goes deep enough to reward thos e who run Kubernetes in production.\n\n## Bio\n\nNilesh Mahajan is a found ing engineer and Head of Infrastructure at Onehouse.ai\, where he builds t he systems that run managed data platforms inside customers' own cloud acc ounts across AWS\, GCP\, and Azure — the deep infrastructure problem at the center of this talk.\n\nHe owns infrastructure\, security\, and compli ance for large-scale data workloads\, including running Spark and other st ateful distributed systems in untrusted\, multi-tenant environments.\n\nBe fore Onehouse\, Nilesh spent years building infrastructure at Uber\, and a longside engineers who built it at LinkedIn and eBay. The platforms they r elied on inside those companies kept data and compute close\, under their own control.\n\nHe now also leads Onehouse's AI initiatives focused on mak ing data platforms run efficiently with less operational overhead.\n\nLink edIn: https://www.linkedin.com/in/nkmahajan/\nX: https://x.com/nilesh_maha jan\n\n## \nhttps://gamma.app/docs/Breaking-the-Trilemma-7y63de2pzr4vepx\n \n## {Add the link to 2-min elevator pitch video}\n\n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T133255Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/breaking-the-trilemma- serverless-data-platform-in-your-own-cloud-account-RoeFyYpY3jR1Uvw8QfUMoh BEGIN:VALARM ACTION:display DESCRIPTION:Breaking the trilemma: serverless data platform in your own cl oud account in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Keynote: Rethinking data architecture for the agent era DTSTART:20260731T043500Z DTEND:20260731T051500Z DTSTAMP:20260728T181111Z UID:session/6r1fLzoEajgJDrJiUZ2Dis@hasgeek.com SEQUENCE:7 CREATED:20260722T104414Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T133652Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Keynote: Rethinking data architecture for the agent era in Aud itorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Full refresh to incremental: rebuilding denormalization for report ing at Razorpay scale DTSTART:20260731T052000Z DTEND:20260731T055000Z DTSTAMP:20260728T181111Z UID:session/N5BdPNrLWDWa7GP2ZcVmqA@hasgeek.com SEQUENCE:17 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T102047Z DESCRIPTION:Abstract / Session Description:\n At Razorpay\, our lake house platform ingests over 6 billion events daily and powers a reporting platform that generates close to a million reports every month. As scale g rew\, full-refresh denormalization became unsustainable: joins across 10-3 0 entities consumed heavy compute\, report freshness lagged by up to 48 ho urs\, and highly mutating datasets made derived tables expensive to keep c orrect.\n\n In this talk\, we’ll share how we moved from full refreshes to an incremental serving model for reporting. Instead of rebuilding tabl es\, we identify impacted records\, use secondary indexes to locate affect ed entities\, and traverse relationships across facts and dimensions using graphs to propagate changes.\n\n We’ll cover the hard cases this fram ework handles\, including dimension-side updates\, backdated foreign-key c hanges\, and out-of-order events. We’ll also discuss how Apache Iceberg features such as table layout optimization\, metadata pruning\, merge tuni ng\, and compaction strategies helped make this practical on our lakehouse platform.\n\n This reduced denormalization compute cost and generation time by over 85%\, while significantly improving freshness and scalability .\n\nWho Is This For?\nThis session is for data platform engineers\, lakeh ouse architects\, analytics infrastructure teams\, and backend engineers w ho build or operate reporting\, CDC\, denormalization\, or derived-data sy stems over large\, highly mutating datasets.\nIt will be especially useful for teams that have outgrown full-refresh pipelines\, are struggling with freshness and recomputation cost\, or are trying to serve reporting workl oads directly from a lakehouse architecture.\n\nKey Takeaways\nWhy full-re fresh denormalization breaks down at reporting scale.\nHow to identify and process only impacted records instead of rebuilding derived tables.\nHow secondary indexes and entity-relationship traversal can support incrementa l propagation.\nHow to handle dimension-side updates\, backdated foreign-k ey changes\, and out-of-order events.\nWhat storage and execution tradeoff s matter when serving reporting workloads from a lakehouse.\nHow Razorpay reduced denormalization compute cost and generation time by over 85%.\n\nD raft slides - [Link](https://drive.google.com/file/d/1w8VmYShWGjaldf33umuh gfBXHu83qM7w/view?usp=sharing)\n\nBio:\nAmit Prabhu\,\nStaff Software Engi neer\, Razorpay\n[Linkedin](https://www.linkedin.com/in/amit-prabhu-179070 a5/)\n\n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T140341Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/full-refresh-to-increm ental-rebuilding-denormalization-for-reporting-at-razorpay-scale-N5BdPNrLW DWa7GP2ZcVmqA BEGIN:VALARM ACTION:display DESCRIPTION:Full refresh to incremental: rebuilding denormalization for re porting at Razorpay scale in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sponsored talk: AIOps: leveraging AI for software incidents DTSTART:20260731T052000Z DTEND:20260731T055000Z DTSTAMP:20260728T181111Z UID:session/KdT1rLrCgRKxBeiBjWSnt2@hasgeek.com SEQUENCE:15 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T102246Z DESCRIPTION:## Description\n\nDuring production incidents or on-call sched ules with a barrage of alerts\, engineers must sift through hundreds or th ousands of services\, code changes\, metrics data points\, logs\, and trac es to reason about the issue and find the root cause.\n\nThis talk explore s how an AI SRE agent inside the Jira Service Management product tackles t his needle-in-a-haystack problem and assists engineers using context graph s\, service graphs\, code intelligence\, observability triangulation\, and agentic workflows.\n\nI'll walk through five building blocks:\n\n1. Dynam ic Runbooks\n2. Grounding using Teamwork Graph and Service Graphs\n3. Arch itecture for Context Engineering\n4. Agent Orchestration\n5. Change/Code I ntelligence\n\nThese building blocks help solve this problem effectively a nd can be applied to similar reasoning problems.\n\nI'll close with real-w orld challenges\, coverage and accuracy metrics\, and lessons learned.\n\n ## Takeaways\n\n1. **Context engineering and search space reduction techni ques** \n Half the battle is getting the right context. Gathering relev ant context using organization-level knowledge graphs such as Teamwork Gra ph\, knowledge bases\, and code intelligence helps the agent significantly reduce the search space for a problem.\n\n2. **Agentic architecture for r easoning** \n Learn how to build agents that reason from first principl es using dynamic plans and a layered architecture consisting of harnesses\ , skills\, tools/actions\, and backend systems to power long-running workf lows.\n\n## Who Should Attend\n\nSoftware engineers and SREs looking to le verage AI for post-deployment alerts and incident response.\n\n## Bio\nSat hish Kumar is a Senior Principal Engineer at Atlassian\, building incident and alert intelligence capabilities in the Jira Service Management produc t.\n\nLinkedIn: https://www.linkedin.com/in/sathish-kumar-588a0452/\n\nX: https://x.com/sathish316\n\n## Draft Slides\n\nWill open up access later - https://docs.google.com/presentation/d/1s_N-ljHt3jDqNXctS2cL03DTX5UTZQ2U2 2cLWplnMuE/edit?usp=sharing\n\n## Elevator Pitch Video\n\nFull video from the Atlassian Unleash conference: https://www.loom.com/share/39d2491795df4 5bd992f606e226628f9 GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T134101Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/aiops-leveraging-ai-fo r-software-incidents-KdT1rLrCgRKxBeiBjWSnt2 BEGIN:VALARM ACTION:display DESCRIPTION:Sponsored talk: AIOps: leveraging AI for software incidents in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Morning break DTSTART:20260731T055000Z DTEND:20260731T061500Z DTSTAMP:20260728T181111Z UID:session/KyBarhh4TP9VB7ECZ8fEQa@hasgeek.com SEQUENCE:12 CREATED:20260722T103209Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260724T112533Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Morning break in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Morning break DTSTART:20260731T055000Z DTEND:20260731T062500Z DTSTAMP:20260728T181111Z UID:session/U8eWnjZuKzoHVhC3gdeeTC@hasgeek.com SEQUENCE:11 CREATED:20260722T103317Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T134113Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Morning break in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Beyond GPUs: cutting ML inference costs by 10× without sacrificin g latency DTSTART:20260731T061500Z DTEND:20260731T064500Z DTSTAMP:20260728T181111Z UID:session/LfYsB18ddYb6BTtcVqpzXz@hasgeek.com SEQUENCE:15 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T102757Z DESCRIPTION:Inference cost-to-serve is usually treated as a fixed tax: the model needs a GPU\, the GPU costs what it costs\, and the bill scales wit h traffic. It isn't fixed. For a large class of production models — embe ddings\, CNNs\, classic CV and NLP — quantization plus graph fusion turn s that GPU tax into a variable you control\, cutting cost-to-serve by ~10 × at the same latency\, throughput\, and accuracy envelope.\n\nThis is th e operational story of doing that across a 40+ model fleet at IDfy\, where we did these migration in a repeatable way\, enough times to become a pla ybook: calibrate → quantize → fuse → validate → canary. I'll work it through one concrete example — a high-volume face-match model moved f rom GPU to BF16 inference on Intel CPUs via OpenVINO\, dropping a 1-RPS po d cost by 10X. I'll then show two more migrations from the fleet\, includi ng one that broke in production and the telemetry that caught it in shadow mode before a single user was affected.\n\nThe savings are the headline ( ~₹50 lakh/year and climbing)\, but the second-order wins were just as im pactful:\n\nGPU capacity reclaimed for the workloads that genuinely need i t\nEnabling cutting edge models for single tenant customers without GPU ac cess with a lower TCO\nA smaller\, burstier CPU footprint you can serve fr om cheaper spot pools with far more elastic autoscaling.\n\nThis is not a "CPU beats GPU" pitch. It's about matching each workload to the cheapest h ardware that still meets its accuracy and latency budget — and the engin eering discipline that tells you\, before production does\, which models s urvive the move and which won't.\n\n## Takeaways\n\nA cost-vs-latency deci sion matrix for GPU vs. quantized-CPU inference — and the one metric tha t actually predicts whether a migration survives production (it isn't aver age latency).\nThe three quantization failure modes we see most often — tail blur\, silent FP32 fallback\, and calibration drift — and the speci fic observability signal that catches each before users do.\nThe business case beyond the raw bill: how cheaper cost-to-serve makes single-tenant\, per-client deployments economically viable (low TCO\, compliance-friendly) and frees GPU capacity for cheaper spot-based autoscaling on the workload s that truly need it.\n\n## Audience\n\nProduction ML and AI engineers\, p latform and infra teams\, and engineering leaders who own inference cost-t o-serve at scale\, plus solution architects who deploy models into client environments\, where deployment model and TCO are part of the solution\n\n Bio\nVivek Kalyanarangan is Sr. Technical Architect\, AI at IDfy\, where a 20-person team operates 40+ production ML models across biometric authent ication\, document recognition and OCR\, fraud detection\, and large-scale NLP. He has 13+ years across analytics\, big data\, and deep learning. Au thor of Quantization and Fast Inference (Manning\, MEAP 2026) and the free CodeCamp course LLMs from Scratch. Contributor to open-source ML and publi shed papers. GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T141607Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/beyond-gpus-cutting-ml -inference-costs-by-10x-without-sacrificing-latency-LfYsB18ddYb6BTtcVqpzXz BEGIN:VALARM ACTION:display DESCRIPTION:Beyond GPUs: cutting ML inference costs by 10× without sacrif icing latency in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sponsored talk: Don’t block AI: handling sensitive data and DPDP while preserving context DTSTART:20260731T062500Z DTEND:20260731T065500Z DTSTAMP:20260728T181111Z UID:session/X9LqRmATU61cvZCTWBuzLg@hasgeek.com SEQUENCE:12 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T104117Z DESCRIPTION:Description\nEnterprise AI is forcing organizations to rethink one of the most fundamental assumptions in data security.\nFor decades\, security strategies have focused on protecting data where it is stored. En cryption\, tokenization\, access controls\, and governance were designed f or applications that queried databases and presented information to users. Generative AI changes that model completely.\nToday's AI systems reason o ver customer conversations\, documents\, support tickets\, medical records \, contracts\, emails\, and enterprise knowledge bases as the very data th at contains the highest concentration of sensitive personal information. O rganizations are now caught between two competing priorities: providing AI with enough context to deliver meaningful outcomes while complying with r egulations like India's Digital Personal Data Protection (DPDP) Act.\nMost enterprises respond in one of two ways. They either expose sensitive data to AI models\, increasing privacy and compliance risks\, or aggressively redact and mask data\, significantly reducing AI quality.\nThis talk expla ins why both approaches are incomplete.\nUsing real enterprise AI architec tures including Retrieval-Augmented Generation (RAG)\, AI copilots\, and a gentic workflows\, the session introduces Runtime Data Control\, an archit ectural pattern that enables AI systems to reason over privacy-safe\, sema ntically preserved representations of sensitive data while revealing ident ities only when required through policy-driven runtime access controls.\nR ather than treating privacy as a compliance checkpoint that slows AI adopt ion\, attendees will see how modern data architecture can enable organizat ions to securely leverage AI while maintaining governance\, preserving con text\, and aligning with regulations such as the DPDP Act.\nThe session co ncludes with a broader perspective: enterprise AI requires us to move beyo nd protecting where data is stored and start governing how data is used.\n \nTakeaways\nAttendees will leave with:\nA practical framework for buildin g enterprise AI without exposing sensitive customer data.\nWhy traditional masking and redaction reduce AI effectiveness and what works better.\nArc hitectural patterns for securing RAG systems\, AI copilots\, and agentic A I workflows.\nHow Runtime Data Control enables AI innovation while strengt hening privacy and governance.\nPractical guidance for aligning enterprise AI architectures with DPDP requirements without compromising AI performan ce.\n\n\nWho Should Attend?\nCTOs and Chief Architects\nAI and ML Engineer ing Leaders\nPlatform and Data Engineering Teams\nSecurity and Privacy Arc hitects\nEnterprise Architects\nCISOs and Data Protection Leaders\nEnginee ring teams building RAG\, Copilots\, or Agentic AI applications\n\nBio\nKa lpesh Jajoo is a Solutions Architect\, APAC at Skyflow\, with more than 20 years of experience helping enterprises build secure\, scalable\, and com pliant technology platforms.\nHe specializes in Data Privacy\, Data Protec tion\, Cloud Infrastructure\, Digital Native Applications\, and Enterprise AI architectures. At Skyflow\, Kalpesh works closely with customers acros s the APAC region to solve complex challenges around protecting sensitive data while enabling AI adoption and regulatory compliance.\nPrior to Skyfl ow\, Kalpesh held engineering and architecture leadership roles at Bayer\, Vodafone\, and Cisco\, delivering large-scale infrastructure and digital transformation initiatives. He holds a Master's degree in Computer Network s from the UK and maintains certifications across Cisco\, Microsoft\, VMwa re\, and Red Hat.\nLinkedIn: https://www.linkedin.com/in/kalpeshjajoo/\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T141558Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/dont-block-ai-handling -sensitive-data-and-dpdp-while-preserving-context-X9LqRmATU61cvZCTWBuzLg BEGIN:VALARM ACTION:display DESCRIPTION:Sponsored talk: Don’t block AI: handling sensitive data and DPDP while preserving context in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:How to Engineer Agent Quality Loops (White-Box) for Reliable AI Ag ents? DTSTART:20260731T063000Z DTEND:20260731T073000Z DTSTAMP:20260728T181111Z UID:session/TB9GpAQiLVxHFdc5U9tNen@hasgeek.com SEQUENCE:3 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,Birds of Feather (BOF) session CREATED:20260728T061007Z DESCRIPTION:# Description\n\nAgent systems work through loops: plan\, call tools\, inspect results\, recover\, and decide when to stop. As these sys tems move beyond demos\, the harder question is not whether they can compl ete a task\, but whether the path was reliable\, efficient\, grounded\, an d worth its cost and what should change when it was not.\n\nThis Birds of a Feather session will examine how to evaluate and improve these agent qua lity loops. When an agent gets stuck\, retries a tool\, wastes tokens\, ca lls the wrong interface\, uses weak evidence\, or returns a plausible but incomplete answer\, how do we determine whether the problem lies in the mo del\, prompt\, tool contract\, context\, orchestration\, or evaluation its elf?\n\nThe panel will establish a shared frame for traces\, logs\, dashbo ards\, and evals: what each can prove and what level of rigor is useful. T he panel will then drive the discussion through concrete failure patterns and turn-level signals\, while participants share their own experiences\, approaches\, and open questions.\n\nTogether\, the room will examine how t o distinguish useful work from wasted turns and token costs\, which signal s lead to meaningful engineering decisions\, and how findings should flow through the quality loop:\n\nObserve the turns → evaluate the path and o utcome → classify the failure → change the system → measure again\n\ nThe goal is to build a shared map of what agent quality loops should cont ain\, what teams are building today\, and which patterns should become com mon practice for production agent systems.\n\n## Discussion Outline\n- Wha t should a useful agent quality view show at the turn level?\n- What is th e difference between agent traces\, logs\, and evals?\n- How should teams measure wasted tokens\, unnecessary turns\, retries\, and avoidable latenc y?\n- How do we distinguish model failures from tool\, context\, schema\, orchestration\, or evaluation failures?\n- Which path-level (agent traject ory) signals reveal real quality problems\, and which are merely interesti ng metrics?\n- How should findings feed back into tooling\, benchmarks\, e valuations\, and release decisions?\n- What level of rigor is appropriate before an agent moves from experimentation to a production workflow?\n\n## Takeaways\n\n- A practical model for evaluating both an agent’s answer and the path it took.\n- A clearer way to turn trace\, cost\, and evaluati on evidence into engineering improvements.\n\n## Who Should Attend?\n\n- A I and platform engineers building production agent systems\n- Technical le ads designing and operating agent quality loops\n- Leaders responsible for agent evaluation\, observability\, cost\, or reliability\n\n\n## Panel\n* *Prateek Mandloi\, Principal Engineer at Atlassian** works on agentic AI s ystems\, enterprise work graphs\, CLI and tooling interfaces\, and benchma rk-driven quality loops for AI agents. Prateek has architected large scale distributed systems and is applying alot of learnings into engineer AI na tive tools and infrastructure at scale. He previously hosted a Birds of a Feather session at The Fifth Elephant 2024 on enterprise data lifecycle an d AI analytics.\n\n**Sunny R Gupta\, Head of Engineering and India Site Le ad at Metaforms**\, an AI platform building agentic workflows for market r esearch agencies. Based in Bengaluru\, he has architected and operated lar ge-scale systems handling billions of events per day across data infrastru cture\, streaming\, and consumer products. Sunny also runs _TeamShiksha_ a nd _Impromptu Meetups_\, mentoring engineers and hosting community learnin g sessions on distributed systems\, data platforms\, and developer product ivity. He brings operability practices from real-world systems\, with a bi as toward first-principles thinking under constraints.\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T061018Z LOCATION:Birds of Feather (BOF) - NIMHANS Convention Centre\nBengaluru\,\n IN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/how-to-engineer-agent- quality-loops-white-box-for-reliable-ai-agents-TB9GpAQiLVxHFdc5U9tNen BEGIN:VALARM ACTION:display DESCRIPTION:How to Engineer Agent Quality Loops (White-Box) for Reliable A I Agents? in Birds of Feather (BOF) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:High concurrency & low latency serving on Apache Iceberg DTSTART:20260731T065000Z DTEND:20260731T072000Z DTSTAMP:20260728T181111Z UID:session/9QySbVSRqXZgji78Z5DtCU@hasgeek.com SEQUENCE:13 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T104508Z DESCRIPTION:## Description\nEveryone is putting their data into Apache Ice berg\; almost no one is serving sub-second queries directly from it. Once data lands in Iceberg\, a familiar question arises: how do you power real- time experiences without duplicating it in yet another serving system? Thi s challenge is especially sharp in observability workloads like RUM\, clic kstream\, and APM\, and in customer-facing analytics dashboards\, where th e cost of low latency often leads to pipeline sprawl. Data teams want stro ng isolation and fast queries\, but not a web of ETL jobs and redundant da ta copies.\n\nIn this talk\, I'll share how we at StarTree built a fast qu ery layer for Apache Iceberg using Apache Pinot\, and what it takes to mak e Iceberg behave like a low-latency analytics store at scale. I'll walk th rough a Kafka to Iceberg to Pinot/StarTree architecture where Iceberg stay s the single source of truth while Pinot powers production queries. The se ssion covers the key design decisions: Pinot indexing and pruning for sele ctive reads\, parallel prefetching of Iceberg blocks over S3\, and trading off local vs remote storage for cost and latency. I'll close with benchma rk results on roughly 1 TB across realistic query shapes and real-world ex amples\, both internal customer metrics exploration and external product a nalytics\, all without extra ETL or duplicate data.\n\n## Takeaways\n1. A practical blueprint for adding a low-latency serving layer on top of Apach e Iceberg without data duplication or additional ETL pipelines.\n2. A refe rence Kafka to Iceberg to Pinot/StarTree architecture with concrete design choices and benchmark data (QPS and latency on roughly 1 TB) for filters and aggregates across primitive and complex types.\n\n## Who Should Attend \nData engineers\, platform engineers\, and backend engineers building rea l-time or user-facing analytics products who are evaluating or already usi ng Apache Iceberg as a data lakehouse foundation.\n\n## Bio\nVivek Sinha i s a Product Manager at StarTree\, where he leads the charter to extend Apa che Pinot's query engine to run directly on Apache Iceberg\, enabling high concurrency\, low latency analytics without data duplication. With over a decade of experience across database systems\, ETL\, data lakes\, and OLA P\, from founding engineer to PM at Hevo Data\, he has shipped data infras tructure products at Fortune 500 scale across batch and real-time processi ng. He also leads the AI Initiative charter for the Ingestion and Applicat ions layer on StarTree Pinot\, shaping how AI-native workloads are ingeste d and served at scale.\n\n[Link to the Draft Slides](https://docs.google.c om/presentation/d/1fUfAlhQYjzok40tlN4GMLArKFlS17tE0n28HEcge8MY/edit?usp=sh aring) GEO:12.9432663;77.5964535 LAST-MODIFIED:20260724T112548Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/high-concurrency-low-l atency-serving-on-apache-iceberg-9QySbVSRqXZgji78Z5DtCU BEGIN:VALARM ACTION:display DESCRIPTION:High concurrency & low latency serving on Apache Iceberg in Au ditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Diagnosing data pipeline failures with AI agents: from research to production\, and the open challenges DTSTART:20260731T070000Z DTEND:20260731T073000Z DTSTAMP:20260728T181111Z UID:session/ApHkEvJtWKdSuVEDwzsf4g@hasgeek.com SEQUENCE:16 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T105616Z DESCRIPTION:Modern enterprise data platforms rely on complex data pipeline s for data transformation and integration\, with thousands running every d ay to move data across systems. When a data pipeline run fails\, the error you see is usually not the root cause. Failures surface far from where th ey are born\, and the component that throws the error is rarely the one at fault\; so diagnosing them means reasoning across logs\, pipeline structu re\, transformation logic\, schema\, and other observability data to find where the failure actually originated\, not where it surfaced. Most observ ability tooling just hands you raw logs\, metrics\, and piles of metadata\ , and leaves the reasoning to you. We built LLM agents that do this reason ing automatically — classify the failure\, trace the causal chain back t o the true root cause\, and propose grounded remediation and shipped them into a production data platform(IBM Watsonx.data integration)\, where they now run against live pipelines. \n\nThis talk is the honest arc of that j ourney: how the agents work\, and the decisions that mattered far more tha n model choice — log and pipeline preprocessing\, reasoning rigor\, hand ling knowledge gaps\, hallucination control\, cost optimization\, and huma n-in-the-loop hooks. We'll cover how we evaluated the system early on with limited data\, and how we measure the real-world impact of these agents. We'll also look at skill-based agents for cases where the agent has direct environment access\, and the new challenges that opens up. And we'll delv e into the problem we're actively tackling now: how do you know the agent' s diagnosis was actually right\, at scale\, when there's no labelled data for every failure? How do you reliably improve the agent over time?\n\nTak eaways:\n1. A generalized diagnosis-agent architecture for data pipelines: classify\, trace the causal chain (symptom vs. cause)\, recommend. \n2. W hat turns a demo into a trusted dataops system: tooling\, preprocessing\, reasoning rigor\, hallucination control\, cost optimization\, and keeping a human in the loop\, all of which matter more than the base model.\n3. Wh y evaluating diagnosis agents is hard: failure-complexity tiers\, no groun d-truth labels\, and scoring reasoning paths rather than answers. Their ev aluation remains a genuinely open problem.\n\nBenefits to the Ecosystem:\n AI agents are being incorporated into data platforms everywhere to diagnos e and remediate pipeline failures\, but little has been said publicly abou t what makes it trustworthy in production or about the uncomfortable fact that almost no one can rigorously measure whether their diagnosis agent is correct across complex scenarios\, let alone improve it reliably over tim e. This talk shares those lessons and challenges.\n\nBio:\nRajmohan C is a Senior Research Engineer at IBM Software Innovation Labs(formerly IBM Res earch)\, Bengaluru\, with over a decade of applied-research experience in Data & AI and a consistent track record of taking innovations from researc h into products with real-world impact. While his past work spans data lin eage and provenance\, data quality\, and GenAI for tabular-data tasks\, he currently works on Agentic DataOps\, specializing in AI-driven diagnosis of data pipelines. He has several peer-reviewed papers and patents to his name. He has contributed to multiple enterprise-scale product GAs over the course of his career.\n\nDraft slides:\nhttps://ibm.box.com/s/369ltgi260e y9t9ck4p4tc364785up5g\n\nElevator pitch:\nTo be added later\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T142018Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/diagnosing-data-pipeli ne-failures-with-ai-agents-from-research-to-production-and-the-open-challe nges-ApHkEvJtWKdSuVEDwzsf4g BEGIN:VALARM ACTION:display DESCRIPTION:Diagnosing data pipeline failures with AI agents: from researc h to production\, and the open challenges in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Beyond metadata: building an open context layer for AI DTSTART:20260731T072500Z DTEND:20260731T074000Z DTSTAMP:20260728T181111Z UID:session/P3QF1m8koMfDaVn1c2Xgs5@hasgeek.com SEQUENCE:10 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T104732Z DESCRIPTION:## Describe your session in 2 paragraphs\n\nEveryone has seen how AI transformed software engineering. Tools like Claude Code and Codex work because code comes with rich context—source control\, dependency gr aphs\, reviews\, tests\, and execution history. Enterprise data\, however\ , lacks an equivalent foundation. AI agents are expected to answer questio ns\, generate SQL\, understand business metrics\, and automate workflows w hile operating on fragmented metadata\, inconsistent business definitions\ , and disconnected systems. Bigger models alone don't solve this problem.\ n\nIn this session\, we'll explore the concept of an **Open Context Layer* *—a new architectural layer that gives AI agents a shared understanding of enterprise data through **Context\, Semantics\, and Memory**. Using Ope nMetadata 2.0 as an open-source case study\, we'll walk through how unifie d metadata graphs\, ontology-driven semantics\, and organizational memory can make AI systems more accurate\, explainable\, and reliable. We'll also discuss the open standards behind this architecture\, including knowledge graphs\, MCP\, OpenLineage\, and RDF\, and share practical lessons from b uilding these capabilities in production.\n\n## 1–2 takeaways from your session\n\n* Understand why enterprise AI systems fail even with state-of- the-art LLMs\, and why context—not model size—is becoming the critical differentiator.\n* Learn practical architectural patterns for building an Open Context Layer using metadata\, semantics\, knowledge graphs\, and or ganizational memory that any AI platform can leverage.\n\n## Which audienc es will benefit?\n\nThis session is intended for:\n\n* Data Engineers\n* D ata Platform Engineers\n* AI/ML Engineers\n* Analytics Engineers\n* Data A rchitects\n* Engineering Managers and Technical Leaders building AI-powere d data platforms\n\n## Speaker bio\n\nMayur Singal is a Lead Integrations Engineer at Collate and an active contributor to the OpenMetadata project. He works on metadata ingestion\, lineage extraction\, and integrations ac ross more than 130 data platforms\, helping organizations build AI-ready d ata ecosystems. His work focuses on open standards\, metadata systems\, an d building the infrastructure that enables reliable AI applications over e nterprise data.\nLinkedin: https://www.linkedin.com/in/mayur-singal-477a76 141/\n\n## Draft slides\nhttps://docs.google.com/presentation/d/1vSnNV6Maq EDjx9kE5uBvCy8zuBatTyIDVZAjove7RoA/edit?slide=id.g3e26a7a2d57_0_7013#slide =id.g3e26a7a2d57_0_7013\n\n\n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260724T112551Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/beyond-metadata-buildi ng-an-open-context-layer-for-ai-P3QF1m8koMfDaVn1c2Xgs5 BEGIN:VALARM ACTION:display DESCRIPTION:Beyond metadata: building an open context layer for AI in Audi torium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:From Figma Node to Production Code: How we Taught Agents to build UI in PhonePe DTSTART:20260731T073500Z DTEND:20260731T075000Z DTSTAMP:20260728T181111Z UID:session/VopWRf3k3ABjs5DvFsPNxc@hasgeek.com SEQUENCE:3 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,15 mins talk CREATED:20260723T134217Z DESCRIPTION:Every design system has the same expensive seam: the handoff f rom Figma to code. A designer ships a pixel-perfect screen\, and an engine er spends hours inspecting layout\, colors\, spacing\, and variants\, then hand-writing it into the codebase over and over\, across every platform t he design system targets. At PhonePe we built FigGen\, an agentic pipeline that turns a selected Figma node into production-ready code for Mint - ou r 150+ component design system. It generates the individual components and then composes them into full screens. But the interesting part isn't "LLM writes UI code." It's everything we had to build around the model to make the output trustworthy: a deterministic component classifier\, a five-lay er architecture contract enforced by a code-review agent\, a compile-time token validator\, and an MCP layer that bridges Figma and our component ca talog.\n\nThis talk walks through the real architecture and the engineerin g decisions that made it reliable. We will show how a conductor agent orch estrates a four-stage pipeline: design analysis\, implementation\, code re view\, finalize - and how the same backbone scales up from a single compon ent to a whole screen by mapping Figma layers onto catalog components. Why we deliberately split analysis and implementation into separate agents to clear accumulated Figma API context before any code is generated. We'll d ig into the parts that actually move accuracy - a decision tree that class ifies every design into one of four component tiers (each mapping to a spe cific code pattern)\, per-variant token resolution tables\, a token valida tor that we compile against the real design-system library to catch invali d tokens before generation\, and BLOCKER gates that stop the pipeline when generated code hardcodes a color or skips an enum variant. Every run is i solated\, versioned\, and instrumented so the agent's work is inspectable and auditable.\n\n### Takeaways:\n- A reusable blueprint for verifiable co de-generation agents: deterministic classification\, an enforced architect ure contract\, and compile-time validation as a feedback loop - patterns t hat apply to any structured output agent\, not just UI code.\n- Why contex t engineering (splitting agents to clear context\, clean handoff packets) and "fail-closed" gates matter more than prompt cleverness when you need a n agent to produce code your team will actually merge.\n\n### Bio:\n**Lalc hand Rathore** is an Architect and Member of Technical Staff at PhonePe. H e specializes in driving cross-platform application stability and performa nce at scale. Currently\, Lalchand is focused on building AI-driven soluti ons that enhance both platform development workflows and end-user experien ces.\n\n**Omkar Sabade** is a Software Engineer and Member of Technical St aff at PhonePe. Working with the payment app core team\, he specializes in iOS application stability and performance at scale. Omkar also focuses on integrating AI capabilities that enhance platform development experiences .\n\n### Who should attend:\nEngineers and platform teams building product ion agents that generate code or other structured artifacts\, design-syste m and frontend-infrastructure owners\, and anyone wiring up MCP servers an d multi-agent workflows who needs the output to be reliable enough to ship .\n\nDraft slides (PDF/PPT\, comments access):\nhttps://drive.google.com/f ile/d/1eJPCEGdwBznlZGPJuyGW4o4vaLhJWqjv/view?usp=sharing GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T135844Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/from-figma-node-to-pro duction-code-how-we-taught-agents-to-build-ui-in-phonepe-VopWRf3k3ABjs5DvF sPNxc BEGIN:VALARM ACTION:display DESCRIPTION:From Figma Node to Production Code: How we Taught Agents to bu ild UI in PhonePe in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch Break DTSTART:20260731T074000Z DTEND:20260731T084000Z DTSTAMP:20260728T181111Z UID:session/Nfnj42LGuJfhsiiDyWjHz5@hasgeek.com SEQUENCE:12 CREATED:20260722T104803Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260727T151516Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Lunch Break in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:From PDF to SQL DTSTART:20260731T075500Z DTEND:20260731T081000Z DTSTAMP:20260728T181111Z UID:session/SPQiyKuu7W9yQTgSx5e2ZK@hasgeek.com SEQUENCE:5 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,15 mins talk CREATED:20260723T134307Z DESCRIPTION:# Fifth Elephant 2026 Submission\n\n ## Description\n\nMost “ask my documents” systems stop at retrieval: upload PDFs\, create emb eddings\, and search for relevant chunks. That works when the user wants s urrounding context\, but it starts breaking down when the user needs numb ers for a report\, dashboard\, analysis\, or presentation. In Indian mutua l fund reports\, the useful answers are usually structured facts: scheme returns\, AUM\, portfolio holdings\, sector allocations\, risk metrics\, b enchmark comparisons\, and month-on-month changes. For these workflows\, t he durable output is not a vector. It is a queryable domain model.\n\nThi s session shares lessons from building a production pipeline that converts Indian mutual fund PDFs\, Excel holdings files\, and visual financial rep orts into structured tables. The system renders pages\, handles OCR and vi sual-table ambiguity\, runs a two-pass LLM workflow for raw capture and sc hema normalization\, writes Parquet outputs\, benchmarks model choices\, t racks extraction cost\, and exposes the final data through OrcaSheets for plain-English and SQL-style querying. I will cover the parts that worked\, and the parts that were painful: dense holdings pages\, page chunking\, i nconsistent AMC formats\, hallucinated rows\, schema drift\, cost control\ , human review loops\, and why converting unstructured data into a known queryable format can be more valuable than repeatedly searching it.\n\n # # 1-2 Takeaways\n\n 1. For numeric insight\, structured extraction often beats vector retrieval. Vectors help find context\, but SQL-like tables ar e better for aggregation\, comparison\, auditability\,\n and repeatabl e answers.\n\n 2. Domain-specific document conversion is hard\, but highl y rewarding. Once messy reports become a known schema\, query cost drops\, outputs become more predictable\, and humans can review rows instead of r eading entire documents again.\n\n ## Audience\n This session will be us eful for:\n\n - Data engineers building pipelines from PDFs\, Excel files \, reports\, or semi-structured business documents.\n - AI engineers work ing on LLM extraction\, OCR\, document understanding\, evaluation\, and co st control.\n - Analytics and BI teams who need reliable numbers from uns tructured sources for dashboards\, reports\, and presentations.\n - Found ers and product engineers building AI-native data products where plain-Eng lish querying must produce predictable\, auditable answers.\n - Teams eva luating when to use vector databases\, Elasticsearch\, SQL\, Parquet\, or structured domain models for document-heavy workflows.\n\n ## Bio\n\n I a m Navdeep\, co-founder/operator building OrcaSheets\, a data product focus ed on making operational and business data queryable through plain English and structured workflows. Our work sits at the intersection of data engin eering\, AI extraction\, and analytics infrastructure: converting messy re al-world inputs such as PDFs\, Excel files\, APIs\, and event streams into reliable data models that teams can query\, review\, and use in reports o r dashboards.\n\nFor this project\, We worked on converting Indian mutual fund reports and holdings data into structured\, queryable tables\, dealin g with visual PDFs\, OCR issues\, LLM benchmarking\, schema normalization\ , Parquet storage\, and cost-aware production workflows.\n\n ## Draft Sli des Link\n[From PDFs to SQL: Turning Indian Mutual Fund Reports into Query able Data](https://drive.google.com/file/d/17kI4MLSSAEId3i33rIW_53U3nqd3ko Rk/view?usp=sharing)\n\n\n\n \n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T135849Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/from-pdf-to-sql-SPQiyK uu7W9yQTgSx5e2ZK BEGIN:VALARM ACTION:display DESCRIPTION:From PDF to SQL in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch Break DTSTART:20260731T081000Z DTEND:20260731T091000Z DTSTAMP:20260728T181111Z UID:session/LRYUz3J8P6m5hRUtc1k3EU@hasgeek.com SEQUENCE:14 CREATED:20260722T110658Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260727T151510Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Lunch Break in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Agentic AI: Path to Production - A Birds of a Feather Discussion DTSTART:20260731T083000Z DTEND:20260731T093000Z DTSTAMP:20260728T181111Z UID:session/MSDjcvWt6ZFjfrCSDDMuDq@hasgeek.com SEQUENCE:7 CREATED:20260728T061131Z DESCRIPTION:## Abstract\n\nBuilding an agent that works in a demo is no lo nger the hard part. The challenge begins when the same system must operate reliably in production\, handling evolving requirements\, multiple agents \, long-running workflows\, changing prompts\, evaluation\, latency constr aints\, and operational failures. Every team encounters different problems \, yet many of the underlying engineering questions are remarkably similar .\n\nThis Birds of a Feather session brings together practitioners who are designing\, deploying\, or operating agentic systems to exchange experien ces from their journey to production. Rather than focusing on frameworks o r model capabilities\, the discussion will centre on architecture\, design decisions\, evaluation strategies\, operational trade-offs\, and the prod uction failures that reshaped those systems. The objective is to surface r ecurring patterns\, identify emerging best practices\, and collectively un derstand what it takes to build agentic systems that remain reliable\, mai ntainable\, and adaptable as they evolve.\n\n## Key Questions That We’ll Explore Together\n\n### Designing Agentic Systems\n\n- What differentiate s a production-ready agentic system from a working prototype?\n- How do yo u decide what should be deterministic code versus agentic reasoning?\n- Wh at architectural patterns have proven effective\, and where have they brok en down?\n- When should a capability be implemented as a tool\, an agent\, or a workflow?\n\n### Consistency & Maintainability\n\n- How do you preve nt prompts\, tools\, schemas\, and business rules from drifting over time? \n- Who owns system state? The application\, the agent\, or a combination of both?\n- How do you build reusable agent components that can evolve acr oss multiple projects and use cases?\n- What practices have improved maint ainability as systems grow in complexity?\n\n### Evaluation & Reliability\ n\n- How do you define and measure success for inherently probabilistic sy stems?\n- What evaluation strategies have proven effective in detecting re gressions before deployment?\n- Which failure modes have been the hardest to identify and debug?\n- How much of your evaluation pipeline is automate d versus human-driven?\n\n### Operating Agentic Systems\n\n- What does obs ervability look like for multi-agent systems?\n- Which metrics\, traces\, and logs have been most valuable in production?\n- How do you diagnose fai lures across multiple interacting agents?\n- What governance and guardrail s have become essential in production deployments?\n\n### Performance & Op timisation\n\n- Where do you see the biggest trade-offs between quality\, latency\, and cost?\n- Which optimization techniques have delivered the gr eatest practical impact?\n- How do you decide when additional reasoning or context is worth the extra latency and cost?\n- What architectural change s have significantly improved scalability or operational efficiency?\n\n## # Lessons from the Field\n\n- What surprised you most after moving from pr ototype to production?\n- What production incident fundamentally changed y our system design?\n- If you were designing your agentic platform again to day\, what would you do differently?\n- What emerging practices do you bel ieve will become standard for production-grade agentic systems?\n\n## Pros pective Participants\n\nThis session is intended for AI engineers\, soluti on architects\, software engineers\, technical leads\, platform engineers\ , and product teams who are designing\, deploying\, or operating agentic s ystems beyond the prototype stage. Whether you’re building internal copi lots\, enterprise automation\, AI products\, or multi-agent platforms\, yo u’ll have the opportunity to exchange experiences with peers tackling si milar production challenges.\n\nI would also encourage participants to com e with one production lesson to share\, whether it’s a design decision t hat worked\, a failure that reshaped their architecture\, or an evaluation practice that caught a critical regression. That simple expectation often leads to far richer discussions than a typical Q&A.\n\n## Key Participant s\n\nTBD - WIP\n``` GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T061249Z LOCATION:Birds of Feather (BOF) - NIMHANS Convention Centre\nBengaluru\,\n IN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/agentic-ai-path-to-pro duction-a-birds-of-a-feather-discussion-MSDjcvWt6ZFjfrCSDDMuDq BEGIN:VALARM ACTION:display DESCRIPTION:Agentic AI: Path to Production - A Birds of a Feather Discussi on in Birds of Feather (BOF) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Discover globally\, materialize locally: building a governed cross -domain data sharing platform DTSTART:20260731T084000Z DTEND:20260731T091000Z DTSTAMP:20260728T181111Z UID:session/2BKLqjxtHMvuxYGzYPUMiG@hasgeek.com SEQUENCE:7 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T105017Z DESCRIPTION:Cross-business-unit data sharing usually starts with good inte ntions and quickly turns into ticket-driven exports\, undocumented copies\ , governance bottlenecks\, and growing compliance risk. At InMobi\, multip le business units operate independent lakehouses\, catalogs\, and data eng ineering organizations. Combining data across these domains creates signif icant business value\, but traditional approaches either require centraliz ed warehouses that undermine domain ownership or ad-hoc data movement that creates governance\, operational\, and compliance challenges.\n\nTo addre ss this\, we built DataBridge\, a governed cross-business-unit data sharin g platform based on a simple principle: Discover Globally\, Materialize Lo cally.\n\nDataBridge enables governed exploration through a unified metada ta catalog and federated query layer built on Trino\, Polaris\, and OpenMe tadata. Teams can discover and analyze datasets in place using short-lived credentials\, policy-enforced access controls\, and bounded exploratory q ueries—without moving data or giving up ownership of their domains. Only when a use case demonstrates repeatable value does the consumer request m aterialization into their own environment\, where governance\, legal\, com pliance\, and lifecycle management processes are applied.\n\nIn this talk\ , we'll share the architectural principles behind the platform\, the decis ion framework used to choose between federation and materialization\, and the lessons learned from operating governed cross-domain data products at scale across both Delta and Iceberg ecosystems.\n\nWe’ll dive into the m etadata-consistency challenges that emerge when data movement is replaced by metadata movement\, including catalog synchronization\, UniForm adoptio n\, schema evolution\, and operating a federated query layer across Delta and Iceberg ecosystems. We'll also discuss how incremental materialization pipelines identify and propagate only changed partitions instead of repea tedly copying entire datasets\, enabling scalable cross-domain data produc ts while preserving ownership boundaries.\n\nFinally\, we’ll share the p roduction bottlenecks that emerged while operating the platform at scale a nd the patterns that kept it reliable without sacrificing governance or do main ownership.\n\nKey Takeaways\n1. Separating Discovery from Data Moveme nt\n\nLearn how governed\, in-place discovery enables teams to validate bu siness value before creating long-lived data copies. We'll share the pract ical signals we use—query frequency\, latency requirements\, cost-per-sc an\, freshness expectations\, and compliance considerations—to decide wh en a consumer should continue federating versus materialize data into thei r own domain.\n\n2. Operating Delta and Iceberg Under a Shared Data Plane\ n\nUnderstand the operational realities of running Delta and Iceberg side by side through a shared catalog and federated query layer. We'll cover me tadata synchronization strategies\, UniForm adoption lessons\, schema evol ution challenges\, cross-domain access control\, and the production issues that emerge when multiple lakehouse technologies must behave like a singl e platform.\n\n3. Building Incremental Materialization at Scale\n\nSee how metadata-driven change detection\, partition-level incremental processing \, and consumer-driven materialization enable scalable cross-domain data p roducts without creating uncontrolled data sprawl. We'll discuss the trade offs\, failure modes\, and operational lessons from running these workflow s in production.\n\nAudience\n\nThis session is aimed at Data Engineers\, Data Platform Engineers\, Data Architects\, and Engineering Leaders buildi ng Data Mesh\, Data Product\, or federated analytics platforms. It will be particularly valuable for teams dealing with cross-domain data sharing\, governance\, metadata management\, and lakehouse architectures built on te chnologies such as Iceberg\, Trino\, Polaris\, OpenMetadata\, and Delta La ke.\n\nBio\n \nShivam Gupta is a Staff Engineer at InMobi\, architect of D SP\, InMobi's real-time bidding platform\, & lead of the materialization p latform of the DataBridge data mesh. https://www.linkedin.com/in/shivam-gu pta-29979834/\n\nKuldip Puri Tejaswi is a Senior Data Engineer on InMobi's central reporting platform team\, leading the query federation layer of t he DataBridge data mesh.\nhttps://www.linkedin.com/in/kuldippuritejaswi/\n \nDeep Patel is a Staff Engineer 2 at InMobi\, lead architect of the DataB ridge data mesh & 1Weather\, InMobi's Climate Intelligence Platform.\n \n\ n\nSlides Link -> https://docs.google.com/presentation/d/1OiFhrpu87Gq6eJkJ n1tNVdWLhqfmlZi1H4QwDwCXV-M/edit?usp=sharing\n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260724T112556Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/discover-globally-mate rialize-locally-building-a-governed-cross-domain-data-sharing-platform-2BK LqjxtHMvuxYGzYPUMiG BEGIN:VALARM ACTION:display DESCRIPTION:Discover globally\, materialize locally: building a governed c ross-domain data sharing platform in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:When There's No Unit Test for "Good": A Maker-Checker Loop for Sub jective AI Output DTSTART:20260731T091000Z DTEND:20260731T094000Z DTSTAMP:20260728T181111Z UID:session/AGFQBThNpAEnoUbPhMyVLw@hasgeek.com SEQUENCE:11 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T110748Z DESCRIPTION:BRIEF DESCRIPTION:\n-\n**The problem:** When an AI agent write s code\, you can test whether the code works. But when an agent makes a ch art\, how do you test whether it's any good? A chart can be technically co rrect and still fail to get its point across. "Good" depends on the audien ce and the decision they need to make - there's no test that returns true or false. \n\nThis talk is about a different approach: two AI agents worki ng as a pair. One makes the chart\, the other reviews it\, and they go bac k and forth - make\, critique\, revise - until it's good enough. Both shar e the same idea of what makes a chart good\; what differs is the angle the y come at it from. I'll use charts as the running example because they hav e a rare advantage - the audience can look at the screen and judge for the mselves\, live\, whether the agents got it right.\n\n**The key insight:** a creator agent can write code to produce a chart\, but it never sees the chart its code produced. It works in code\; the chart is a picture. So it misses things you'd only catch by looking - heavy gridlines fighting the d ata\, labels printed to three decimal places\, a cluttered layout\, an imp ortant number left unhighlighted. It's blind a second way too: it already knows what it meant to say\, so it can't read the result with fresh eyes t he way a real audience would. \n\n**The solution:** The reviewer agent is built to defeat both blind spots - it renders the chart and looks\, and it reviews without being told what the maker intended\, so it reacts like a real viewer. That difference in what each agent can see and know is what s tops the two from just agreeing with each other. \n\nI'll close by demoing a maker and a reviewer agent working a real chart end to end - including a moment where one of them gets it wrong - and show how the back-and-forth becomes a signal for improving the system over time.\n\nKEY TAKEAWAYS:\n- \n1. A practical way to think about quality for AI outputs that have no ri ght answer - starting with the difference between problems you can only se e by looking and problems in whether the audience will understand.\n2. How to design a maker-reviewer agent pair so that giving each agent different things to see and know keeps them from rubber-stamping each other's work. \n\nWHO IT'S FOR:\n-\n- Engineers and teams building AI that produces work judged by taste rather than correctness - charts\, slides\, documents\, d esigns\, writing.\n- Anyone wrestling with how to measure quality when "be tter" is partly subjective.\n- Anyone who wants to learn how to design clo sed loops for AI capability improvement\n- (No data-visualization backgrou nd needed\; charts are just the example)\n\nBIO:\n-\nHi\, I'm Vikram Nayak \, founder of ChartBoss. We help compabies benchmark and improve the capab ilities of their AI systems at chart and slide generation. \n\nWebsite: ht tps://www.chartboss.ai\nLinkedIn: https://www.linkedin.com/in/vikramnayak8 5\nX: https://x.com/vikramnayak85\n\nI've spent 18 years in BI and Analyti cs\, with the last 6 of those dedicated to data visualizationn and data pr oducts. \n\nI've helped companies like Trendlyne (a stock-market analytics platform with 1M+ users) and Stylumia (AI fashion intelligence)\, and hav e run data-visualization workshops at Delhivery\, Stylumia\, and NSRCEL @ IIM Bangalore. This talk is the engineering behind the product\, not a pro duct walkthrough.\n\n**Slides**: https://docs.google.com/presentation/d/1Z 0VXolOvT1JqAwwIF5BZqvk859iNGcZZ/\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T135904Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/when-theres-no-unit-te st-for-good-a-maker-checker-loop-for-subjective-ai-output-AGFQBThNpAEnoUbP hMyVLw BEGIN:VALARM ACTION:display DESCRIPTION:When There's No Unit Test for "Good": A Maker-Checker Loop for Subjective AI Output in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Your voice agent is (probably) doomed\, OR how not to fall victim to outdated voice agent playbooks DTSTART:20260731T091500Z DTEND:20260731T094500Z DTSTAMP:20260728T181111Z UID:session/At5m8gBU6R5o1Wb3LVudjE@hasgeek.com SEQUENCE:15 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T105810Z DESCRIPTION:Googly Bhai was busy this IPL season. He live-streamed to an * *audience of 300\,000 every day\, with 1.5 million minutes of watch time a nd 1.1 million concurrent viewers at peak**. Hundreds of other streamers c alled him onto their streams to discuss live scores\, gossip\, and make pr edictions ([see him live jamming with another streamer](https://youtu.be/c acqJSyiyZM)). He switched to Aussie and British accents mid-conversation a t the audience's request and **effortlessly switched between 80+ languages \, including Bengali\, Telugu\, and Punjabi**. \n\nGoogly Bhai\, despite t he name\, is not a small-time Mumbai goon. He's the **world's first real-t ime AI sports streaming companion** - a sarcastic\, street-smart half-monk ey who loves Virat (and lets it show). He reacts to live match events\, ru ns trivia mid-stream\, reads audience chats in real time\, and pulls up li ve stats - all at roughly 2-second latency.\n\nWhen we started building th e **AI voice companion for live sports entertainment for our 5M daily user s on Dream11** in 2025\, conventional wisdom asked us to treat it as an ex tension of LLM chatbots. Unfortunately\, we followed this advice\, and it cost us **5 months of wasted time building and discarding 8 prototypes and rewriting our entire codebase three times** before arriving at something that worked. The insights we earned are surprising\, non-trivial\, and unt il now\, insider knowledge. Through this talk\, I want to share them with everyone. \n\nThe chatbot extension school of thought assumes that buildin g a voice agent means adding plumbing over an LLM-powered chatbot. **Just add an STT and a TTS\, sandwich the LLM in the middle\, and you've built a voice agent\, right? Right? WRONG!** Nothing could be further from the tr uth. Voice deserves its own native components:\n\n- **Full duplex speech-t o-speech models** - not an LLM sandwiched between an STT and a TTS\n- A ** network protocol built for voice** - not TCP\, which was designed to bring your mail\n- A way to **inject functionality without blocking or overpowe ring the conversation** - not tools and MCPs\, which work alright for chat bots but fail spectacularly for voice agents.\n\nIn addition\, I'll share how to think about **scaling concurrency in a world where everyone is GPU- bound**.\n\n**All of these lessons will be corroborated by real numbers in production that we will share !** \n\n# What I hope the audience will tak e away\n\nA battle-tested playbook they can take back to their desk and us e on Monday. A playbook that doesn't exist in the open right now. Four spe cific lessons:\n\n1. Cascaded STT→LLM→TTS pipelines are broken for con versations. Latency adds up and **conversational context decays at compone nt boundaries**.\n\n2. WebSockets/REST APIs on TCP will kill real-time voi ce experience because **TCP values correctness more than latency**. WebRTC on UDP was built for multimedia communication and prioritizes latency ove r correctness. \n\n3. Tool calls create dead air and increase cognitive ov erload for models. People can wait 10 seconds for an *answer* but not for a *response*. There is a difference between latency and **perceived latenc y** with voice agents that is important to get right. \n\n4. The size of y our model providers matters. Bigger providers can offer pay-as-you-go pric ing for higher scales since they can aggregate demand. **Small labs hit GP U provisioning walls fast**. And if you're thinking about **self-hosting** \, I will break your heart with facts in this talk.\n\n\n# Qui Bono? Who w ill benefit?\n\nEngineers building voice agents or real-time AI applicatio ns where conversation quality is paramount. Product and engineering leader s deciding architecture for voice-first products. \n\n\n# Dhruv Nigam - Sp eaker Bio\n\n**Staff ML Engineer\, Dream11**\n\n[Website](https://dhruvnig am93.github.io/) · [LinkedIn](https://www.linkedin.com/in/dhruv-nigam-525 31176/) · [GitHub](https://github.com/dhruvnigam93) · [Twitter/X](https: //twitter.com/druubeey)\n\n---\n\nDhruv Nigam is a Staff ML Engineer at Dr eam11\, India's largest fantasy sports platform. At Dream11\, he built [LU MOS](https://arxiv.org/abs/2512.08957)\, a 200M-parameter foundation model trained on 1.7 trillion tokens that replaced 50+ task-specific models and lifted DAU by 3% overnight. Most recently\, his team launched [World's fi rst live sports AI influencer](https://www.youtube.com/watch?v=zkMaDzV08-I )\, a multi-agent voice system that has reached over a million users. Befo re Dream11\, he worked as a quantitative analyst at an investment bank\, b uilding trading algorithms on billions in capital.\n\nHe has presented at KDD Barcelona\, Swiggy\, and MumbaiPy among other venues and writes about ML engineering at [ML Trenches](https://mltrenches.substack.com/). His int erests lie in LLM post-training\, inference optimization\, and agentic sys tems.\n\n[Other Talks](https://dhruvnigam93.github.io/talks) · [Writing]( https://dhruvnigam93.github.io/writing) · [Publications](https://dhruvnig am93.github.io/writing#publications)\n\n\n# References\n\nSlides(WIP) - \n https://docs.google.com/presentation/d/1XFxKHhwdDqJx7EIQW53EIVxL84DEn1AOgg UDI3yJpUY/edit?usp=sharing\n\nOfficial Dream11 blog about the AI companion - \nhttps://medium.com/dreamlockerroom/how-we-built-the-worlds-first-real- time-ai-sports-streaming-companion-a42273de1fca\n\nPersonal blog on lesson s learned building voice agents- \nhttps://mltrenches.substack.com/p/what- i-learned-building-voice-agents\n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T142051Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/your-voice-agent-is-pr obably-doomed-or-how-not-to-fall-victim-to-outdated-voice-agent-playbooks- At5m8gBU6R5o1Wb3LVudjE BEGIN:VALARM ACTION:display DESCRIPTION:Your voice agent is (probably) doomed\, OR how not to fall vic tim to outdated voice agent playbooks in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Beyond SWE-bench: How do we evaluate AI agents for real-world work flows? DTSTART:20260731T093000Z DTEND:20260731T103000Z DTSTAMP:20260728T181111Z UID:session/7q6DqXcYWrwTUf3JPsBdKV@hasgeek.com SEQUENCE:5 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,Birds of Feather (BOF) session CREATED:20260728T061054Z DESCRIPTION:**Track:** Track 2 – Building & Implementing AI Tools & Agen ts in Production \n**Format:** Birds of a Feather (BoF) Session \n\n## T itle\nHow do we evaluate AI agents for domain-specific\, real-world workfl ows?\n\n## Description\nAI agents deployed in real-world products do more than generate a response. They interpret user goals\, make decisions\, ret rieve information\, use tools\, follow policies\, take actions\, recover f rom failures\, and determine when to involve a human.\n\nEvaluating such s ystems is challenging because there is often no public benchmark or univer sally correct answer for the workflow being automated. What constitutes a successful agent depends heavily on the product\, domain\, and intended us er outcome.\n\nFor example\, an agent handling a customer-support request may need to resolve the underlying problem\, follow organisational policie s\, minimise user effort\, avoid unsafe actions\, and escalate appropriate ly. Similar challenges arise in incident response\, IT operations\, data q uality\, finance operations\, procurement\, compliance\, and other domain- specific workflows.\n\nIn this BoF session\, we will discuss how AI practi tioners\, product teams\, and domain experts can jointly translate the int ended outcome of an agentic feature into meaningful evaluation criteria.\n \nThis includes identifying domain-specific metrics\, determining whether to evaluate only the final outcome or also the agent’s decisions and act ions\, and defining acceptable trade-offs between quality\, user experienc e\, reliability\, latency\, and cost.\n\nWe will also explore how teams ca n build representative evaluation datasets when no ready-made benchmark ex ists. Potential sources include production traces and logs\, historical ca ses\, domain-expert-authored scenarios\, observed failures\, and synthetic ally generated examples.\n\nAn important part of the discussion will be ho w offline evaluation results can be validated against online product\, use r\, operational\, and business metrics after deployment.\n\nThe goal is to exchange practical approaches for answering a central question:\n> _How d o we know that an agentic product feature is genuinely effective\, reliabl e\, and valuable in its intended domain?_\n\nThe discussion will focus on evaluation design\, domain reasoning\, datasets\, metrics\, and product ou tcomes rather than any particular framework or implementation stack.\n\n## Discussion Outline\n**Defining the outcome**\nWhat user\, operational\, o r business problem is the agent expected to solve? What would constitute s uccess\, partial success\, or failure in that domain?\n\n**Defining domain -specific metrics**\nHow do teams derive metrics for:\n\n- Resolution qual ity\n- Task completion\n- Policy adherence\n- Human handoff quality\n- Saf ety & Latency\n- Other product or domain-specific requirements\n\n**Evalua ting outcomes and agent behaviour**\nIs evaluating the final result suffic ient\, or should teams also examine:\n\n- The agent’s decisions and inte rmediate actions\n- Tool selection and tool usage\n- Adherence to constrai nts and policies\n- Recovery from errors or incomplete information\n- Esca lation and human-handoff behaviour\n- Whether the chosen path was efficien t and appropriate\n\n**Building representative evaluation datasets**\nHow can teams combine:\n\n- Production traces and logs\n- Historical cases\n- Domain-expert-authored scenarios\n- Observed failure modes\n- Synthetic da ta\n\nHow can this be done without losing realism\, introducing dataset le akage\, or overfitting to previously observed cases?\n\n**Connecting offli ne and online evaluation**\nHow do we determine whether improvements in of fline evaluation actually correlate with:\n\n- Better user experience and Improved business outcomes\n- Higher task-resolution rates\n- Safer agent behaviour\n- Reduced latency or cost\n\n**Choosing evaluation and judging methods**\n- Which criteria can be evaluated deterministically?\n- Which r equire human or domain-expert review?\n- Where can LLM-based judges be use ful\, and how should they be calibrated or validated before being trusted? \n\n## Key Takeaways\nParticipants should leave with:\n\n- A clearer appro ach for translating product and domain goals into measurable evaluation cr iteria.\n- Ideas for evaluating both the outcome of an agentic workflow an d the decisions or actions taken to reach it.\n- Practical strategies for creating evaluation datasets when no public benchmark or labelled dataset exists.\n- A better understanding of how production traces\, domain expert ise\, synthetic scenarios\, edge cases\, and failure analysis can compleme nt one another.\n- Approaches for connecting offline evaluation scores wit h online product\, user\, operational\, and business metrics.\n- A framewo rk for deciding which aspects of evaluation should use deterministic check s\, human review\, domain experts\, or LLM-based judges.\n- A shared under standing of which parts of agent evaluation may be reusable across domains and which must remain specific to the product and workflow.\n\n## Who Sho uld Attend?\nThis session is intended for cross-functional teams building and evaluating agentic systems\, particularly:\n\n- Applied AI\, ML\, data science\, and evaluation practitioners designing metrics\, datasets\, rub rics\, graders\, and evaluation processes.\n- Product managers and product leaders defining agentic features and determining whether those features deliver meaningful user and business value.\n- Engineering and platform te ams operating production agents and investigating reliability\, latency\, cost\, safety\, observability\, and failure modes.\n- Domain experts and o perations teams whose knowledge is required to define correct behaviour\, important edge cases\, organisational policies\, and acceptable outcomes.\ n- Open-source contributors and researchers building agent-evaluation tool s\, datasets\, test environments\, observability systems\, or reusable eva luation methodologies.\n\nParticipants do not need to work in a particular domain. Examples may include customer support\, incident response\, IT op erations\, CRM\, compliance\, procurement\, finance operations\, data work flows\, developer tools\, and other products in which agents perform multi -step tasks.\n\n## Suggested Reading and References\n\n1. **A Survey on Ev aluation of LLM-based Agents**\nA broad overview of agent evaluation cover ing planning\, tool use\, interactive environments\, application-specific benchmarks\, safety\, robustness\, and evaluation using humans or language models (https://arxiv.org/abs/2503.16416)\n\n2. **AgentBoard: An Analytic al Evaluation Board of Multi-turn LLM Agents**\nIntroduces fine-grained pr ogress metrics for evaluating what an agent accomplishes during multi-step execution\, rather than relying only on final task success (https://arxiv .org/abs/2401.13178)\n\n3. **τ-bench: A Benchmark for Tool-Agent-User Int eraction in Real-World Domains**\nEvaluates agents that interact with user s\, use domain tools\, modify environment state\, and comply with policies in realistic workflows (https://openreview.net/forum?id=roNSXZpUDN)\n\n4. **WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?**\nDemonstrates how a domain-specific evaluation environment can be built around realistic enterprise knowledge-work tasks (https://arxiv.org /abs/2403.07718)\n\n5. **CRMArena-Pro: Holistic Assessment of LLM Agents A cross Diverse Business Scenarios and Interactions**\nExplores realistic ev aluation of agents across sales\, service\, and business workflows\, inclu ding multi-turn interactions\, confidentiality requirements\, and expert-v alidated tasks (https://arxiv.org/abs/2505.18878)\n\n6. **Demystifying Eva ls for AI Agents**\nA practical guide to designing agent evaluation tasks\ , datasets\, graders\, and evaluation processes during product development (https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T061316Z LOCATION:Birds of Feather (BOF) - NIMHANS Convention Centre\nBengaluru\,\n IN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/beyond-swe-bench-how-d o-we-evaluate-ai-agents-for-real-world-workflows-7q6DqXcYWrwTUf3JPsBdKV BEGIN:VALARM ACTION:display DESCRIPTION:Beyond SWE-bench: How do we evaluate AI agents for real-world workflows? in Birds of Feather (BOF) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:TiDB sponsored talk DTSTART:20260731T094500Z DTEND:20260731T101500Z DTSTAMP:20260728T181111Z UID:session/U8R1isNG99RVya6EafdoNs@hasgeek.com SEQUENCE:12 CREATED:20260722T105048Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T142250Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:TiDB sponsored talk in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sponsored talk: The practical guide to reverse-engineering XXL cod ebases with agentic AI DTSTART:20260731T095000Z DTEND:20260731T102000Z DTSTAMP:20260728T181111Z UID:session/3PVP6V8ZR5QXkag5SJgU7b@hasgeek.com SEQUENCE:17 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T110820Z DESCRIPTION:\n## Abstract\nPoint an AI coding agent at an unfamiliar codeb ase and ask it to explain the architecture\, and for a few hundred files\, it works beautifully. Point the same agent - with the same well-engineere d prompt - at a ten-thousand-file enterprise Java monolith\, and it quietl y falls apart: the context window fills\, compaction kicks in\, the model starts forgetting details\, and then it starts inventing them. This isn't a prompt-quality problem. It's a physics problem: a fixed-size processing container being asked to hold an unbounded amount of code.\n\nThis talk is a production report on solving it. Working on a real\, ten-thousand-file legacy Java system\, we built a five-layer\, eleven-agent pipeline that re verse-engineers architecture\, domain models\, service boundaries\, and wo rkflows - without ever asking a single agent to read the whole codebase. D eterministic tools (build introspection\, LSP\, tree-sitter) handle struct ure\; LLM agents are reserved exclusively for semantic reasoning\; and a l ightweight orchestrator (the Conductor) coordinates specialised agents tra cking only the pipeline state\, never code content. The result: 10\,000+ c lasses indexed\, 200+ living documents\, 400+ database tables mapped\, 140 + API endpoints catalogued\, and documentation that regenerates incrementa lly as the code changes.\n\nI'll walk through the architecture layer by la yer - Scout\, Cartographers\, Specialists\, Surveyors\, Chronicler\, Condu ctor - and distill it into four transferable principles of context enginee ring: progressive disclosure\, deterministic grounding\, role isolation\, and subagent isolation. I'll also share what broke along the way: a symbol -indexing layer that had to be rebuilt for speed\, the point where the ful l pipeline stopped being worth its cost\, and why decoupled agent layers l et us fix any one of them without touching the rest.\n\n### Target audienc e\nEngineers and architects using AI agents on real codebases - especially anyone who has hit the wall where a coding agent starts hallucinating on a large\, unfamiliar system - plus teams designing multi-agent pipelines f or structural or high-volume analysis work.\n\n### Key takeaways\n\n- A co ncrete architecture pattern for multi-agent pipelines that stay within con text limits by design\, not by luck\n- A working example of separating det erministic (tool) work from cognitive (LLM) work at scale\n- Four reusable principles of context engineering\, and the production lessons that shape d them\n- How to tell when the full pipeline is worth building - and when a thinner approach is the better call\n\n## Draft slides\nhttps://drive.go ogle.com/file/d/18J_UrBgILfz_bhZ06rCyZCF5wi_RTSEn/view?usp=drive_link\n\n# # About the speaker\n- **Bio**: Harshad has over 16 years of experience bu ilding scalable software\, distributed systems\, and data products. With a background deeply rooted in core software engineering and agile practices \, he is currently helping large enterprises on AI enablement and legacy s ystem modernization—leveraging agentic AI to safely navigate\, decode\, and transform complex codebases.\n- **Company**: Sahaj Software\n- **Job r ole**: Solutions Consultant\n- **Links**:\n - [LinkedIn](https://www.lin kedin.com/in/harshad-nawathe-76762618/)\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T142326Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/the-practical-guide-to -reverse-engineering-xxl-codebases-with-agentic-ai-3PVP6V8ZR5QXkag5SJgU7b BEGIN:VALARM ACTION:display DESCRIPTION:Sponsored talk: The practical guide to reverse-engineering XXL codebases with agentic AI in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Evening break DTSTART:20260731T101500Z DTEND:20260731T110000Z DTSTAMP:20260728T181111Z UID:session/DP8Ps838dxtrVduJf1myYS@hasgeek.com SEQUENCE:9 CREATED:20260722T164856Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T140002Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Evening break in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Break DTSTART:20260731T102000Z DTEND:20260731T105000Z DTSTAMP:20260728T181111Z UID:session/CMjCUWF4oEPRofCfq8hgSz@hasgeek.com SEQUENCE:11 CREATED:20260722T164949Z GEO:12.9432663;77.5964535 LAST-MODIFIED:20260724T112734Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Break in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Preserving Human Intuition in AI Development DTSTART:20260731T104000Z DTEND:20260731T113500Z DTSTAMP:20260728T181111Z UID:session/QShSkz5U4rgfaZVrHefwLC@hasgeek.com SEQUENCE:7 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,Birds of Feather (BOF) session CREATED:20260728T060933Z DESCRIPTION:Three major studies from Anthropic\, METR\, and MIT recently r evealed an uncomfortable reality about AI-assisted development. Their data shows that AI tools can lower skill assessment scores\, increase actual t ask completion times\, and significantly reduce neural connectivity during creative problem-solving. This is not a simple case of developers becomin g lazy. Instead\, widespread AI usage actively disrupts the cognitive feed back loops essential for building deep technical expertise. This Birds of a Feather session focuses on how we can prevent the silent erosion of our baseline intuition\, the fingertip feeling of engineering\, while continui ng to work alongside these models.\n\nThis is a practitioner-led conversat ion rather than a lecture or an anti-AI rant. We will spend the first few minutes framing the problem around personal guardrails like Friction Budge ts and Verification-First Protocols\, then open the floor to the room. The rest of the hour belongs to the community to share real-world war stories \, debate tactics\, and figure out how to maintain our value as engineers in an automated world.\n\n## Takeaways\n- **Practical Cognitive Guardrails :** You will walk away with peer-tested strategies to reintroduce intentio nal cognitive friction into your daily coding loop\, ensuring you stay sha rp while utilizing LLMs.\n\n- **Team Policy Ideas:** Engineering leads wil l gain actionable approaches for structuring team workflows that preserve junior developer growth without sacrificing shipping speed.\n\n## Target A udience\nThis session is highly beneficial for software developers who use tools like GitHub Copilot or Cursor daily and want to ensure they are not outsourcing their critical thinking. It is equally valuable for engineeri ng managers and team leads who are wrestling with AI adoption policies\, c ode quality retention\, and mentoring the next generation of engineers.\n\ n## Speaker Bio\nI’m a Solution Consultant at Sahaj Software in Chennai and Ambassador of Agentic AI Foundation. I’ve been neck-deep in MCP\, AI -Assisted Software Development\, and agentic architectures since before th ey had proper names. I’ve spoken at GitTogether 2025\, The Fifth Elephan t 2025\, and many other conferences across AI-assisted development\, MCPs and privacy-preserving ML. Right now I’m obsessed with: what happens to developer cognition when AI writes most of your code\, why nobody’s buil ding agent runtimes for the browser yet and MCP Security. GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T061302Z LOCATION:Birds of Feather (BOF) - NIMHANS Convention Centre\nBengaluru\,\n IN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/preserving-human-intui tion-in-ai-development-QShSkz5U4rgfaZVrHefwLC BEGIN:VALARM ACTION:display DESCRIPTION:Preserving Human Intuition in AI Development in Birds of Feath er (BOF) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:An Agent That Builds Agents: AI-Powered Recipe Generation for Loca l-First ETL DTSTART:20260731T105000Z DTEND:20260731T112000Z DTSTAMP:20260728T181111Z UID:session/824TVsUtswYZba4njzSmWB@hasgeek.com SEQUENCE:9 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T105541Z DESCRIPTION:## An Agent That Builds Agents: AI-Powered Recipe Generation f or Local-First ETL\n\n## Session Description\n\nMost organisations have ru n Spark or Glue jobs for datasets that comfortably fit in a single machine 's memory. \nFor about five years now\, the real bottleneck has not been i nfrastructure\; it has been tooling. OrcaSheets is a desktop application t hat lets analysts build full ETL workflows locally\, with no cluster provi sioning\, no cloud egress\, and no YAML configs. The remaining challenge w as adoption: our internal DSL for transformation pipelines could not be ex posed directly to users\, and the unlearning curve had to be smooth. So we built an agent that generates agents. \nThe Recipe Engine takes natural l anguage descriptions and produces validated\, repeatable pipelines\, not j ust SQL but complex algorithms like k-means clustering and time-series for ecasting\, all running on a Rust binary with an extremely small footprint. For 90% of real-world workloads\, OrcaSheets is a legitimate alternative to cloud ETL.\n\nThis talk challenges the default assumption that data eng ineering requires distributed systems. We'll show production benchmarks co mparing local execution against equivalent Glue/Spark jobs and demonstrate the recipe generation system end-to-end: from a messy 2GB CSV upload thro ugh AI-generated transformations to a clean\, versioned output with cluste ring\ninsights. We'll dive into the meta-problem of building an agent that understands your internal DSL well enough to generate other agents (recip es)\, the prompt engineering\, validation loops\, and guardrails that make this reliable in production. The audience will leave with a concrete fram ework for deciding when local-first is the right architecture and how an A I-powered recipe engine eliminates the boilerplate that makes ETL expensiv e.\n\n## Takeaways\n\n1. How to build an "agent that creates agents": the architecture of a recipe generation engine that translates natural languag e into validated\, repeatable transformation pipelines using an internal D SL\, including the prompt engineering and validation patterns that make it production-reliable.\n2. A practical framework for when local-first analy tics replaces cloud ETL (Spark/Glue)\, with real cost comparisons showing 10-50x savings and how AI-generated recipes reduce pipeline development ti me from days to minutes while keeping data fully on-premises.\n\n## Target Audience\n\nData engineers exploring alternatives to cloud ETL for mid-si ze datasets\; ML engineers interested in local-first analytical tooling\; platform engineers building internal DSLs or agent-powered developer tools \; anyone building AI agents that generate structured\, validated outputs rather than freeform text.\n\n## Bio\n\nMayur\, CTO at Orcasheets\, Co-fou nder at DataOrc. I build practical\, scale-proof systems\, starting from Q A\, through backend/frontend\, to co-founding one of the fastest analytics engines today.\n\nAt OrcaSheets\, I’m making analytics that run locally on your PC\, handling 100M+ rows in seconds and turning raw data into ins ights\, dashboards\, and decisions.\n\nBefore this\, I co-founded DataOrc\ , a data engineering consultancy that’s delivered for 75+ enterprises\, helping them scale data pipelines\, cut cloud bills by up to 90%\, and st ay production-grade at petabyte scale.\nLinkedin: https://www.linkedin.com /in/mayur-jadhav-1a3a4652/ \nTwitter: https://x.com/mayurJ13\n\n## Draft S lides\nhttps://drive.google.com/file/d/1nefNm1c_c_9LjFfEaLwl7bKVWVY87aRa/v iew?usp=sharing\n\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260724T112722Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/an-agent-that-builds-a gents-ai-powered-recipe-generation-for-local-first-etl-824TVsUtswYZba4njzS mWB BEGIN:VALARM ACTION:display DESCRIPTION:An Agent That Builds Agents: AI-Powered Recipe Generation for Local-First ETL in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Many Narrow Agents\, Not One Big Prompt: Putting AI on the On-Call Rotation DTSTART:20260731T110000Z DTEND:20260731T111500Z DTSTAMP:20260728T181111Z UID:session/6v6Y5awrhpnqvJdw73qrmd@hasgeek.com SEQUENCE:5 CATEGORIES:Track 1 - Data engineering & infrastructure,15 mins talk CREATED:20260723T134540Z DESCRIPTION:Abstract\nThe obvious way to put an AI agent on the on-call ro tation is one large prompt that knows everything. We rejected it for a mul ti-agent architecture: around 18 specialized agents over a large cloud dat abase\, a router that classifies each incident to the right specialist\, a nd a shared library of nearly 300 reusable skills. Diagnosis is automated but mitigation stays human-approved.\n \nThis talk is about the reliabilit y engineering that architecture demanded. Composing agents reintroduces di stributed-systems failure modes - routing loops\, cascading handoffs\, and plausible-but-wrong reasoning that never surface in a demo. We'll walk th e guardrails that contain them: bounded routing\, loop prevention\, ground ing every conclusion in telemetry\, and the human-in-the-loop checkpoints that decide where autonomy stops.\n \nYou'll leave with a framework-agnost ic blueprint for agentic incident response - how to decompose one agent in to a fleet\, and the SRE patterns that keep that fleet reliable under real production load.\n \nFor: SREs\, on-call engineers\, and teams putting ag ents into production incident response.\n\n\nBio:\nHi I'm Hrithik. I work at Microsoft Bengaluru in the Azure SQL team. Linkedin: http://linkedin.co m/in/hrithik-piyush/\n\nSlides Deck: https://drive.google.com/file/d/1OeFT nH5Tds4KJLhN5-P03Y5E0JLFYrs2/view?usp=sharing GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T140058Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/many-narrow-agents-not -one-big-prompt-putting-ai-on-the-on-call-rotation-6v6Y5awrhpnqvJdw73qrmd BEGIN:VALARM ACTION:display DESCRIPTION:Many Narrow Agents\, Not One Big Prompt: Putting AI on the On- Call Rotation in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Vector DBs Are Overrated: Grounding LLM Agents on Master Data With out the Overhead DTSTART:20260731T112000Z DTEND:20260731T113500Z DTSTAMP:20260728T181111Z UID:session/TK5PxU8ftFUEpGyUMTYiXQ@hasgeek.com SEQUENCE:4 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,15 mins talk CREATED:20260723T134709Z DESCRIPTION:Large Language Models are powerful at reasoning and recommenda tion\, but they routinely hallucinate entities that do not exist in an org anization’s proprietary data. The common response is to build a Retrieva l-Augmented Generation (RAG) stack with embeddings\, a vector database\, a nd retrieval infrastructure. However\, many enterprise datasets are not co llections of long documents—they are structured master data such as prod uct catalogs\, audience segments\, taxonomies\, brands\, and reference lis ts containing thousands of short records. In our case\, we needed to groun d an agent’s recommendations in a proprietary catalog while keeping late ncy\, operational complexity\, and deployment footprint low.\n\nThis sessi on presents a practical evaluation of lightweight retrieval approaches for agent grounding without a vector database. We benchmarked BM25 variants\, MiniLM semantic embeddings\, and Model2Vec static embeddings across real- world planning queries and evaluated them not only on recall\, but also on a “trap rate” metric that measures how often retrieval surfaces wrong -but-plausible results that could mislead a downstream LLM. The findings w ere surprising: lightweight static embeddings came within a point of trans former-based semantic retrieval while requiring a fraction of the runtime footprint\, and enriching records with descriptions often increased the nu mber of confusable false positives without significantly improving recall. We will share the evaluation methodology\, results\, and a practical fram ework for selecting retrieval strategies based on how agents consume propr ietary data.\n\n### Key Takeaways\n* You may not need a vector database to ground agents against structured enterprise data. Lightweight approaches such as BM25 and static embeddings can achieve comparable retrieval qualit y with significantly lower operational overhead\, latency\, and deployment complexity.\n* Measure retrieval safety\, not just retrieval accuracy. Im proving recall is only part of the problem\; wrong-but-plausible retrieval results can be more harmful to agent behavior than obvious misses. Unders tanding and measuring these “trap” results leads to more reliable grou nded agents.\n\n### Intended Audience\nThis session will be valuable for:\ n* AI/ML Engineers building agentic applications and LLM-powered workflows \n* Platform and Infrastructure Engineers evaluating RAG architectures\n* Data Scientists and Applied AI practitioners working with proprietary ente rprise data\n* Product Engineers integrating LLMs with catalogs\, taxonomi es\, reference data\, or recommendation systems\n* Engineering leaders loo king for pragmatic alternatives to complex vector database deployments\n\n [Draft slides deck](https://docs.google.com/presentation/d/1owsM6wJxN6y5yX rB8qDTcT8rZ2Rgf9jM58cvupODoPU/edit?usp=sharing) GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T140110Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/vector-dbs-are-overrat ed-grounding-llm-agents-on-master-data-without-the-overhead-TK5PxU8ftFUEpG yUMTYiXQ BEGIN:VALARM ACTION:display DESCRIPTION:Vector DBs Are Overrated: Grounding LLM Agents on Master Data Without the Overhead in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Architecting Observability Platform on S3 + Lambda DTSTART:20260731T112500Z DTEND:20260731T115500Z DTSTAMP:20260728T181111Z UID:session/HQuYijUMsx7iKAb6ev7drB@hasgeek.com SEQUENCE:14 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T105131Z DESCRIPTION:Observability is a Data Engineering problem. Write-heavy\, rea ltime latency\, faster queries are the typical requirements of a Observabi lity system. Traditional observability systems fail to meet the ever-growi ng demand of increased volumes due to cloud deployments\, AI agents speedi ng up the feature and product development. AI agents also changes the quer y patterns which were common in Observability - human-led sparse queries. AI agents can make 100x more queries than human engineers which breaks the assumptions underlying fixed compute and disk based observability archite ctures. This session goes into building blocks required to design an obser vability system on Object storage and Serverless architecture.\n\n## Key T akeaways\n1. An architectural blueprint of an observability system on top of object storage and serverless compute. It includes how to design data f ile formats and index file formats optimized for S3 dynamics.\n2. Utilizin g serverless compute for bursty loads is essential for reducing the idle t ax of always-on clusters.\n\n## Target Audience\nThis session is ideal for \n- Data engineering architects who are building scalable and performant s ystems on object storage.\n- SRE/DevOps engineers navigating cost / operat ional overhead challenges of traditional observability systems\n\n### Bio\ nFounding Engineer at [Oodle AI](https://oodle.ai)\, an observability plat form built on object storage + serverless compute.\n\nPrev: Early (among f ist 30 hires) engineer at Rubrik\n\n- [LinkedIn](https://www.linkedin.com/ in/sujeetgholap/)\n- [Twitter](https://x.com/sujeetgholap)\n- [sujeet.ghol ap@oodle.ai](mailto:sujeet.gholap@oodle.ai)\n\n### Draft slides\n- [Google Slides](https://docs.google.com/presentation/d/1Mm1_y8F9B4EXPgjz8t2CnEcdS QAiQACuG_s0fTLhvE4/edit?usp=sharing)\n\n### Elevator pitch:\nto be added GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T142650Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/architecting-observabi lity-platform-on-s3-lambda-HQuYijUMsx7iKAb6ev7drB BEGIN:VALARM ACTION:display DESCRIPTION:Architecting Observability Platform on S3 + Lambda in Auditori um 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:When the orchestrator wasn’t enough: a production story of build ing reliable multi-agent AI DTSTART:20260731T114000Z DTEND:20260731T121000Z DTSTAMP:20260728T181111Z UID:session/2fmrVeb2NvT1PRaHri3tKz@hasgeek.com SEQUENCE:7 CATEGORIES:Track 2 - Building & implementing AI tools & agents in producti on,30 mins talk CREATED:20260722T111011Z DESCRIPTION:We built a production multi-agent assistant for a complex ente rprise planning workflow where AI assists\, rather than replaces\, human d ecision-making. Users express their intent through natural language\, the system extracts structured semantic information for human review\, and mul tiple specialist agents collaborate to retrieve data\, perform analysis\, and generate recommendations. Our initial architecture followed a familiar orchestrator-driven design. The first production incidents taught us some thing unexpected: every individual agent behaved correctly\, yet the overa ll system did not. We assumed orchestration would keep multiple agents con sistent. Production taught us that consistency is established before orche stration begins. Ambiguous user requests were interpreted differently acro ss agents\, conversational context caused semantic state to drift\, and in dependently correct outputs became collectively inconsistent. Retrieval op erated on different assumptions\, downstream reasoning diverged\, and resp onse generation was left reconciling conflicting interpretations. We reali zed we were treating consistency as an orchestration problem when it was f undamentally a language understanding problem.\n\nThe solution was an arch itectural pivot inspired by lessons from traditional NLP rather than a mor e sophisticated orchestration. We moved language understanding from an imp lementation detail inside the orchestrator to a shared architectural capab ility that every agent depends on. Intent understanding\, entity grounding and disambiguation and state-transition-aware conversation history/contex t management establish a canonical semantic state before orchestration beg ins. The extracted entities became both the contract between downstream ag ents and the UI through which human planners interact with the system. Rat her than presenting another orchestration framework\, this talk shares the production failures that drove this redesign\, the engineering trade-offs behind it\, and why decades-old NLP concepts including intent\, entities\ , and conversational state - remain one of the strongest foundations for b uilding reliable production multi-agent systems in the era of agentic AI.\ n\n**Key takeaways**\n- Learn why consistency failures in production multi -agent systems are often natural language understanding (NLU) problems\, a nd see how classical NLP concepts can be repurposed as architectural build ing blocks for modern agentic AI.\n- Understand how to establish a shared semantic state that keeps downstream agents consistent\, even across ambig uous and multi-turn conversations.\n\n**Target audience**\n- AI/ML enginee rs building production-grade LLM and multi-agent applications\, aiming at deploying reliable\, scalable AI systems in production.\n- Solution archit ects designing enterprise AI systems and agentic workflows.\n- NLP practit ioners interested in applying classical language understanding techniques to modern agentic architectures. Practitioners looking for pragmatic\, eng ineering-first solutions to production challenges.\n- Developers building conversational assistants and human-in-the-loop decision support systems.\ n\n**Speaker bio**\nDr. Karthika Vijayan is a Solution Consultant at Sahaj Software. She has been conducting research in the field of conversational AI with voice and text data for almost a decade. Her research has been pu blished in several journals and presented at various international confere nces. Prior to joining Sahaj Software\, she worked as a research fellow at the National University of Singapore and at IISc Bangalore. She has done her PhD from IIT Hyderabad.\n \nPrevious talk links:\n1. Fifth El Bangalor e (2025): https://www.youtube.com/live/4Nw89rICm-s (6:09:55 - 6:47:30)\n2. Devday Pune (2025): https://youtu.be/1JSZE-1Zlvg?feature=shared\n3. WiDS (2025): https://www.youtube.com/watch?v=_es29z32_Xs (1:26:40 to 2:02:10)\n 4. Devday Bangalore (2025): https://youtu.be/WQUt5Sjf5U8?feature=shared\n\ nProfile links:\nhttps://scholar.google.com/citations?user=fJp6O0UAAAAJ&hl =en\nhttps://www.linkedin.com/in/karthika-vijayan/\nhttps://www.researchga te.net/profile/Karthika-Vijayan\n\n**Draft Slides Deck** \nhttps://docs.go ogle.com/presentation/d/1RL8O7ZTl2B5DzZOHHcMO4fYPmuEgrFt4eTzyZ7SrSQ8/edit? slide=id.g361d9497010_0_1560#slide=id.g361d9497010_0_1560 GEO:12.9432663;77.5964535 LAST-MODIFIED:20260723T140251Z LOCATION:Auditorium 2 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/when-the-orchestrator- wasnt-enough-a-production-story-of-building-reliable-multi-agent-ai-2fmrVe b2NvT1PRaHri3tKz BEGIN:VALARM ACTION:display DESCRIPTION:When the orchestrator wasn’t enough: a production story of b uilding reliable multi-agent AI in Auditorium 2 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Things We Wish We Knew Before Building Agentic Systems DTSTART:20260731T114500Z DTEND:20260731T124500Z DTSTAMP:20260728T181111Z UID:session/CdecxfMvX3oEZZP6WZB7kY@hasgeek.com SEQUENCE:6 CREATED:20260728T061204Z DESCRIPTION:Everyone is building agents. Most demos look impressive. Produ ction systems tell a different story.\n\nLet’s share the assumptions tha t turned out to be wrong\, the architectural decisions we’d change\, and the engineering lessons that only appear after running agentic systems in the real world.\n\nRather than discussing frameworks or model benchmarks\ , we’ll focus on lessons learned from building and operating agentic sys tems in production or at scale.\n\nTopics may include evaluation\, orchest ration\, multi-agent coordination\, state management\, tool use\, observab ility\, guardrails\, cost\, latency\, and deciding when not to use an agen t.\n\n## Discussion prompts\n- What assumption about agentic systems turne d out to be wrong?\n- What production issue surprised you the most?\n- Whe re did you replace LLM reasoning with deterministic code?\n- When is a mul ti-agent architecture actually worth the complexity?\n- How do you evaluat e and debug agentic systems?\n- What tooling\, observability\, or testing practices proved valuable?\n- Looking back\, what would you design differe ntly today?\n- How have your prompts evolved over time? Do prompts “rot ” as the system grows\, requiring constant maintenance?\n- What techniqu es have helped make prompts more robust and maintainable?\n- What retrieva l strategy worked best in practice? Simple RAG\, hybrid search\, knowledge graphs\, SQL\, hierarchical retrieval\, or something else?\n\n## Expected takeaways\nParticipants should leave with:\n\n- Practical engineering les sons from real-world agentic systems\n- Common architectural pitfalls and ways to avoid them\n- Better intuition for when to use agents and when sim pler approaches are sufficient\n- A broader understanding of production ch allenges such as evaluation\, orchestration\, state management\, observabi lity\, guardrails\, and cost\n\n## Bio\nSujit Kamthe is a Solution Consult ant at Sahaj Software based in Pune\, India. He focuses on data engineerin g\, large scale data processing\, and architecting robust data platforms. He has previously presented at The Fifth Elephant on pragmatic guides to r obust data quality checks\, and his work centers on building enterprise sy stems that are scalable\, governed\, and ready for modern AI consumption. GEO:12.9432663;77.5964535 LAST-MODIFIED:20260728T061333Z LOCATION:Birds of Feather (BOF) - NIMHANS Convention Centre\nBengaluru\,\n IN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/things-we-wish-we-knew -before-building-agentic-systems-CdecxfMvX3oEZZP6WZB7kY BEGIN:VALARM ACTION:display DESCRIPTION:Things We Wish We Knew Before Building Agentic Systems in Bird s of Feather (BOF) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Billboards near a coffee shop: making proximity search fast in Cli ckHouse DTSTART:20260731T120000Z DTEND:20260731T121500Z DTSTAMP:20260728T181111Z UID:session/BQVU7dobeZgmsZQneEyQ6e@hasgeek.com SEQUENCE:12 CATEGORIES:Track 1 - Data engineering & infrastructure,30 mins talk CREATED:20260722T105708Z DESCRIPTION:## Title\n\nBillboards near a coffee shop: making proximity se arch fast in ClickHouse\n\n## Session description\n\nWe build planning sof tware for out-of-home advertising: billboards\, and the places people want to be near. Almost everything comes back to one query\, show me the billb oards near our coffee shops\, and it sounds simpler than it is. There are about three million billboards on one side\, twenty million places on the other\, and "near" is not one number\, it changes with every search. Big o n both sides\, with a radius that keeps moving\, it is the kind of query a n analytics engine is worst at\, and every easy way to speed it up fell ap art.\n\nThis is the story of fixing that one query. We tried the obvious t hings first\, and each one hit a wall. What actually worked was not a fast er query\, it was a different question: we changed what we were asking\, a nd ClickHouse happened to have the one thing that made the new question ch eap. I will walk through what we tried\, why it broke\, the idea that fixe d it\, and the two lessons I would take to any hard join.\n\n## Takeaways\ n\n- Know when\, and how much\, to precompute. The obvious precompute can cost more than it saves.\n- Change the question\, not the query. Often the fix is not a faster query but a different one.\n\n## Who is this for\n\nD ata and platform engineers who serve queries at scale\, anyone doing geosp atial or proximity search\, and anyone who likes a good battle story. A li ttle SQL helps.\n\n## Bio\n\nAmaan is an engineer with four years of exper ience. He started a fullstack engineer\, then moved into large-scale data\ , building and running data workflows on Spark\, Airflow and ClickHouse\, in Scala and Python. He now works on data infrastructure for an ad-tech an alytics platform.\n\nhttps://docs.google.com/presentation/d/1Z6mYKi6feA1fk SjqTOSW4RzUQAjCKQAD/\n GEO:12.9432663;77.5964535 LAST-MODIFIED:20260724T112714Z LOCATION:Auditorium 1 - NIMHANS Convention Centre\nBengaluru\,\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2026/schedule/billboards-near-a-coff ee-shop-making-proximity-search-fast-in-clickhouse-BQVU7dobeZgmsZQneEyQ6e BEGIN:VALARM ACTION:display DESCRIPTION:Billboards near a coffee shop: making proximity search fast in ClickHouse in Auditorium 1 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Closing DTSTART:20260731T121500Z DTEND:20260731T123000Z DTSTAMP:20260728T181111Z UID:session/AUQz1E22h9ykXgqeitkxL3@hasgeek.com SEQUENCE:4 CREATED:20260723T141822Z LAST-MODIFIED:20260723T141848Z LOCATION:Bangalore & Hybrid ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Closing in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT END:VCALENDAR