BEGIN:VCALENDAR VERSION:2.0 PRODID:-//HasGeek//NONSGML Funnel//EN DESCRIPTION:Less hype. More engineering. X-WR-CALDESC:Less hype. More engineering. NAME:The Fifth Elephant 2025 Annual Conference X-WR-CALNAME:The Fifth Elephant 2025 Annual Conference REFRESH-INTERVAL;VALUE=DURATION:PT12H SUMMARY:The Fifth Elephant 2025 Annual Conference TIMEZONE-ID:Asia/Kolkata X-PUBLISHED-TTL:PT12H X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT SUMMARY:Check-in at the Registration Desk DTSTART:20250719T031500Z DTEND:20250719T034500Z DTSTAMP:20260421T080241Z UID:session/KmYwEjZ2QzeuEszFsNkr87@hasgeek.com SEQUENCE:2 CREATED:20250702T163508Z LAST-MODIFIED:20250711T151255Z LOCATION:Bangalore & Hybrid ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Check-in at the Registration Desk in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Introduction | Agentic AI Track DTSTART:20250719T034500Z DTEND:20250719T035500Z DTSTAMP:20260421T080241Z UID:session/ExkVjUfqmhhLDQmyjeypTW@hasgeek.com SEQUENCE:4 CREATED:20250702T163448Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T141451Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Introduction | Agentic AI Track in Agentic AI track (Auditoriu m) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T035500Z DTEND:20250719T040500Z DTSTAMP:20260421T080241Z UID:session/MhVk1pkEWvcPaWnQy4V34i@hasgeek.com SEQUENCE:3 CREATED:20250704T145952Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164142Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:E-commerce infographics at AI speed DTSTART:20250719T040500Z DTEND:20250719T044000Z DTSTAMP:20260421T080241Z UID:session/QiKXejHSBDxbEvUuqHPATz@hasgeek.com SEQUENCE:19 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Visual AI track CREATED:20250702T163755Z DESCRIPTION:## Abstract\nAn Infographic is a product image on Product Disp lay Page (PDP) of an e-commerce website such as target.com\, highlighting key features of the product in the image itself reducing the need to scrol l down to read product specifications. Displaying the highlights of a prod uct directly on images is shown to drive higher demand and increased guest conversion due to speeded up purchase decision process. \n\nInfographic c reation at Target has been completely manual process and suffers from very low coverage for products including first and third party brands. In addi tion\, the design of the infographics is tightly controlled by creative te ams and is extremely subtle\, leading to limitations of existing tools to support any automations within infographics space. \n\nIn this session\, I will talk about a *patent pending* solution to generate infographics at s cale for any brand using design template\, text generation and vision comp uting. Specifically\, I will talk about the design elements within infogra phic template\, GenAI based text generation\, product image segmentation a nd simulation to obtain the optimal product crop for infographics.\n\n## K ey Takeaways\n- Understand the design requirements for a general infograph ic image\n- Look at LLM based feature (text) generation given product desc ription \n- Optimally crop the product image using object segmentation and simulation\n- Understand architecture to handle generic template design\n \n## Audience\n- Creative Designers \n- Image processing / Computer Vision experts\n- Developers / Data Scientists building multi-modal pipelines\n\ n\n## Presentation Outline\n- Introduction *(2 mins)*\n- Design elements of Infographics *(3 mins)*\n - Know about the key design elements \n - Understand key idea that is used in this solution\n- Generating Text with LLM *(3 mins)*\n - Generating Attribute + Run-On \n- Product Image *(12 mins)*\n - Automated segmentation of primary and lifestyle images\n - Optimal crop of product based on simulation\n- Architecture and Process F low *(5 min)*\n - General pipeline to handle text and image processing w ithin generic template\n - Future work\n\n## Bio\nShubham Tripathi is a data science professional with over 10 years of experience in retail and a erospace industry. He holds 4 patents (2 in filing) and many trade secrets within the imaging space. He is currently working as Lead Data Scientist at Target\, developing Vision and GenAI based solutions for digital market ing creative design business. Previously\, he had worked for Boeing R&D wh ere he invented and led the development of many imaging solutions for impr oving cabin experience of passengers. He holds BTech and MS in Computer Sc ience from IIIT Hyderabad. \n\n## Slide Deck\n\n[Infographics-FifthEl-2025 .pptx](https://1drv.ms/p/c/96c1707df00266c5/EXfSCwLmiXlBqcTG5A7l0OABwprik7 KyqgT8T8QuNCTs3A?e=PGyuUe)\n\n## Pre-Talk Reading List\n[some blogs\, conc epts to look through before the talk](https://onedrive.live.com/personal/9 6c1707df00266c5/_layouts/15/Doc.aspx?sourcedoc=%7B671ea118-f318-47f3-adf3- dbe131f959e8%7D&action=default&redeem=aHR0cHM6Ly8xZHJ2Lm1zL3cvYy85NmMxNzA3 ZGYwMDI2NmM1L0VSaWhIbWNZOF9OSHJmUGI0VEg1V2VnQmpRNlBuWDVlQWVnSk5zenNYUjkwbU E_ZT1JM1JjVkk&slrid=dbf9b0a1-10f7-9000-35c7-ba3e10a08717&originalPath=aHR0 cHM6Ly8xZHJ2Lm1zL3cvYy85NmMxNzA3ZGYwMDI2NmM1L0VSaWhIbWNZOF9OSHJmUGI0VEg1V2 VnQmpRNlBuWDVlQWVnSk5zenNYUjkwbUE_cnRpbWU9X1hXNlJoTEIzVWc&CID=50193726-f8c a-4ae1-b6b4-69e5192caa5d&_SRM=0:G:185)\n\n## References\n[Infographic Exam ple @ Target.com](https://target.scene7.com/is/image/Target/GUEST_1b9eb45c -7538-4685-977e-0cde8279570f?wid=1000&hei=1000&qlt=80&fmt=webp) GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T071046Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/infographics-image-gen eration-for-e-commerce-QiKXejHSBDxbEvUuqHPATz BEGIN:VALARM ACTION:display DESCRIPTION:E-commerce infographics at AI speed in Agentic AI track (Audit orium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Introduction | Data Engineering and Infrastructure (DEI) track DTSTART:20250719T041000Z DTEND:20250719T042000Z DTSTAMP:20260421T080241Z UID:session/QxCMyELcfka7mjBmL4CNaF@hasgeek.com SEQUENCE:3 CREATED:20250704T153310Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T145337Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Introduction | Data Engineering and Infrastructure (DEI) track in Data Engineering and Infrastructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T042000Z DTEND:20250719T043000Z DTSTAMP:20260421T080241Z UID:session/UrUH3mvKkpzusi7zP1EctF@hasgeek.com SEQUENCE:2 CREATED:20250704T153323Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164118Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Data Engineering and Infrastructure track (Semin ar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Driving ML experiments without breaking the infra budget DTSTART:20250719T043000Z DTEND:20250719T050500Z DTSTAMP:20260421T080241Z UID:session/DLuYxrTbhWjtWdNrDw3uCp@hasgeek.com SEQUENCE:11 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Data & ML Infrastructure track CREATED:20250704T153216Z DESCRIPTION:At InMobi\, data is foundational to everything we do\, powerin g everything from personalization to predictive modeling across our produc ts and platforms. Our data platform ingests\, processes\, and stores petab ytes of data\, and supports a wide spectrum of users\, from product analys ts to ML engineers. \n\nIn this talk\, I’ll walk through how we’ve arc hitected the platform to enable rapid experimentation for data scientists\ , while keeping infrastructure costs in check\, a balancing act critical t o our success.\n\nWe'll begin with an overview of InMobi’s data ecosyste m and technical stack\, covering our use of distributed storage\, Spark fo r large-scale compute\, and the orchestration tools that bind it all toget her. From there\, I’ll motivate why fast turnaround times for ML experim ents from feature engineering to model training are crucial to InMobi's ap plied science workflows. The need for fast iterations must be met without sacrificing resource efficiency\, especially at our scale.\n\nThis led us to define three core tenets that guide how our platform is designed and op timized:\n\n1. What is stored - minimizing redundant and stale data\, pref erring late materialization and pointer-based joins.\n\n2. What is process ed - structuring compute patterns to limit unnecessary shuffles and redund ant reads.\n\n3. How efficiently we process it - the focus of the rest of this talk\, especially around Spark.\n\nIn particular\, we’ve invested d eeply in instrumentation and observability within Spark. We extended the S park Event Listener interface to extract rich runtime metrics\, configurat ion state\, and query plans. But unlike basic Spark UIs or log aggregators \, our observability stack is not a matter of fact event history. We use i t to surface performance bottlenecks\, suboptimal parallelism\, and other tuning opportunities.\n\nBuilding on this\, we’ve integrated with Vertex AI’s Agents Developer Kit (ADK) to develop a multi-agent recommender sy stem. These agents collaboratively reason over Spark metrics\, source code context\, prior review history\, and active Git branches to suggest tunin g recommendations\, auto-generate pull requests\, and flag regressions. Th e goal is to not just observe but act on inefficiencies.\n\nWe orchestrate this flow periodically\, using job metadata and cost traces to drive down infrastructure waste over time both proactively and as part of postmortem feedback loops. \n\nIf time permits\, we’ll walk through a minimalist d emo of this flow end-to-end.\n\nWe’ll conclude by sharing some key learn ings and outcomes including measurable cost savings\, reduced iteration ti me for data scientists\, and improved visibility across stakeholders. Fina lly\, we’ll look at what’s next: expanding beyond Spark\, generalizing the recommender agent framework\, and making performance tuning collabora tive\, explainable\, and self-correcting by design.\n\n### Bio\nSrikanth S undarrajan is a seasoned architect with over 25 years of industry experien ce\, including more than 15 years specializing in large-scale data process ing and distributed systems. A passionate open-source advocate\, he is a m ember of the Apache Software Foundation and has served on the Project Mana gement Committees (PMC) of several Apache projects. Currently\, he leads p latform initiatives at InMobi Technologies\, driving innovation and scalab ility across their systems.\n\n## Pre-Talk Reading List\n[Some blogs\,conc epts to go through before the talk](https://docs.google.com/document/d/11P ymd_iKw95Gjeny3KyXT4J_ixSLAaIw6Ee1gO9uK1M/edit?tab=t.0#heading=h.5kq2oor9c xdy) GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T070918Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/driving-ml-experimenta tion-without-breaking-the-infra-budget-DLuYxrTbhWjtWdNrDw3uCp BEGIN:VALARM ACTION:display DESCRIPTION:Driving ML experiments without breaking the infra budget in Da ta Engineering and Infrastructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T044000Z DTEND:20250719T045000Z DTSTAMP:20260421T080241Z UID:session/Asm6FxxSSwBQfhWXnEN8ky@hasgeek.com SEQUENCE:3 CREATED:20250702T163546Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250704T150005Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Improving storage observability with time mixers DTSTART:20250719T045000Z DTEND:20250719T052500Z DTSTAMP:20260421T080241Z UID:session/GckQAe76Ds5wWx2shEMBVa@hasgeek.com SEQUENCE:16 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,AI-driven analytics track CREATED:20250702T163650Z DESCRIPTION:As enterprises increasingly rely on Software-Defined Storage (SDS) to manage cloud-scale infrastructure\, ensuring optimal workload pla cement\, capacity planning\, and performance monitoring has become a compl ex challenge. Traditional AI-driven observability solutions for IT infrast ructure monitoring often rely on device-specific data and short-term stati stics\, which restricts their ability to generalize across workloads and s torage environments. Recent advancements in Time Series Foundation Models (TSFM) present a promising alternative by leveraging pretrained representa tions on diverse long-term datasets\, enabling transfer learning across de vices and improving predictive accuracy.\nOur work introduces a TSFM-based approach for AI-driven storage management\, leveraging the TinyTimeMixer (TTM) architecture. Our solution\, in production\, incorporates domain-spe cific features to enhance TSFM pretraining for storage performance forecas ting. Through extensive experiments on real-world FlashSystem performance metrics data\, we demonstrate significant improvements over traditional ma chine learning baselines. These results highlight the potential of TSFM-po wered observability to enhance automated storage management.\n\nIn this ta lk\, we will dive into the steps taken to customize general purpose time s eries foundation models like TTM to solve storage domain performance obse rvability use case in production. \n### Outline\n- TTMs - Brief introducti on\n- Storage domain performance observability\n- Extended-pretraining of TTMs\n- Performance analysis compared to classical forecasting models\n- D eployment status\n\n### Key Takeaways\n- Learn about Time Series Foundati on Models\n- Customization of TSFM\, in-particular TTM\, for storage useca ses\n- Naunces of TTM in production…\n\n### Speaker Bio\nChandramouli Ka manchi is a Research Scientist working at IBM Research Bangalore\, with ov er 6 years of industry experience. \nHe obtained his PhD from IISc Bangalo re. His primary areas of interest are Reinforcement Learning\, Stochastic Approximation\, Time Series Forecasting. His expertise lies in solving rea l world problems through mathematical modeling. Beyond his professional pu rsuits\, he is an amateur chess player\, and enjoys cooking.\n\n### Profil e links\n [https://scholar.google.co.in/citations?user=1QlrvHkAAAAJ&hl=en ]\n### Resources\nTTM - Publication [https://proceedings.neurips.cc/paper_ files/paper/2024/file/874a4d89f2d04b4bcf9a2c19545cf040-Paper-Conference.pd f]\n\nTTM in Storage Insights [https://www.redbooks.ibm.com/redpieces/pdfs /redp5755.pdf]\n\n### Presentation\nSlide deck is available [here](https:/ /docs.google.com/presentation/d/1LK8gZeCRvVudwTnzig96-nCmgqmcPqLL/edit?usp =sharing&ouid=112408808359482737071&rtpof=true&sd=true)\n\n### Pre-Talk Re ading List\n[Some blogs\,concepts to go through before the talk](https://d ocs.google.com/document/d/1u6HqY_KRw8h1qK0ENXxGiRNJ6Df0gc1C34RcuDbK9mw/edi t?tab=t.0#heading=h.5kq2oor9cxdy) GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T071023Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/enhancing-performance- observability-through-tiny-time-mixers-for-storage-domain-GckQAe76Ds5wWx2s hEMBVa BEGIN:VALARM ACTION:display DESCRIPTION:Improving storage observability with time mixers in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Intra action: rethinking UI in the age of AI DTSTART:20250719T045000Z DTEND:20250719T052500Z DTSTAMP:20260421T080241Z UID:session/Hqr1SXVWgHF59Yf93LKpxw@hasgeek.com SEQUENCE:17 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Applied AI Engineering & Agentic AI track CREATED:20250702T163837Z DESCRIPTION:**Intra action: rethinking UI in the age of AI**\n\n*Why butto ns and text boxes wonʼt cut it anymore*\n\n## Context\nFor over half a ce ntury\, user interfaces (UIs) have been built on metaphors borrowed from t he physical world : think desktops\, files\, folders\, buttons\, checkboxe s\, and text fields. These interfaces assumed the *user* *knows* *what* *t hey* *want*\, and the system is a passive responder waiting for explicit i nput. This made sense in an era when computers were deterministic tools : calculators\, word processors\, web browsers : each doing exactly what it was told\, no more\, no less.\n\nWith the rise of Artificial Intelligence especially large language models - computers are no longer passive tools\; they have become *active* *collaborators*.AI can interpret ambiguous inst ructions\, generate new ideas\, anticipate needs\, and adapt behavior in r eal-time.\n\nThis means the classic “one click\, one responseˮ or “on e text box\, one inputˮ paradigm cannot capture the fluidity of these cap abilities.\n\nThe core problem is that our interfaces are stuck in the age of typewriters and control panels — but our technology has leapt into t he age of thinking companions. Itʼs time to close this gap.\n\nTo that en d I propose INTRA ACTION\n\nInspired by concepts from Karen Baradʼs philo sophy: unlike “interactionˮ (two separate entities acting upon each oth er)\, *intra* *action* recognizes that human and AI agency emerge *togethe r*\, co-shaping the outcome in real-time.\n\nIn practical terms\, it means interfaces should be:\n\n* **Context-aware:** Sensing user behavior\, env ironment\, emotion.\n* **Proactive:** Anticipating intent and making sugge stions.\n* **Adaptive:** Evolving based on usage history.\n* **Multimodal: ** Combining voice\, gesture\, text\, visuals seamlessly.\n* **Conversatio nal:** Allowing natural back-and-forth without rigid form-filling.\n\nFor example:\n**Typing versus talking:** People increasingly dictate messages or ask voice assistants instead of typing\, expecting conversational nuanc e.\n\n**Smart homes:** People get frustrated when they must tap five butto ns to adjust lighting instead of saying\, “Set cozy mode.ˮ\n\n## The op portunity\nDesigning for AI means embracing:\n\n* *Co-creation* instead of command-and-control.\n* *Fluid\,* *multimodal* *interfaces* that blend te xt\, voice\, vision\, and gesture.\n* *Interfaces* *that* *dissolve* *into * *context* - less screen\, more situational intelligence.\n* *Ethical* *d esign* that explains how AI thinks and makes decisions.\n\n## The stakes\n If we donʼt rethink UI:\n\n* Users will get overwhelmed by systems that a re simultaneously too smart and too dumb — powerful but locked behind pr imitive controls.\n* AI will make mistakes without human-friendly ways to correct or supervise.\n* Designers will miss opportunities to make technol ogy feel truly human-centered.\n\n## Key takeaways for the audience\n* How Interaction design for AI softwares is diffrent from Interaction Design f or Desktop and Mobile.\n* A frame work for approaching the challenges for designing for a fuzzy\, non binary paradigm of AI and LLMs.\n* New UI elem ents \, Input methods \, Interaction methods and symbol convention to comm unicates with machines .\n\n## Who is the audience for this talk \n* Found ers\n* Product Designers\n* Product Managers\n* Engineers\n* Usability res earchers\n* Tech anthropologist\n\n## About the speaker\nPrashant heads pr oduct management at Jar\, a fintech startup revolutionizing the financial landscape in India. Prashant has deep expertise in product management\, de sign\, and leadership across the consumer internet\, mobile apps\, fintech \, commerce\, and AI domains.\n\nBefore Jar\, he was the co-founder of Shi fu\, an AI company in the smartphone-based predictive intelligence space. Shifu was acquired by Paytm\, where Prashant worked as Vice President of P roduct Management for five years\, managing everything from Recharge & Bil l Payment to DIY Merchant Onboarding\, eCommerce Sellers\, and Messaging a nd Content.\n\nHe is very passionate about AI\, fintech\, and conversation al interfaces\, and is dedicated to designing products for users in Tier 2 and Tier 3 towns of India\, also known as the "Next Billion Users.”\n\n Beyond the professional realm\, Prashant's interests include traveling\, r eading\, and culinary experiments. Currently calling Bangalore home\, Pras hant is surrounded by his tech arsenal\, including two MacBooks\, a pair o f tablets (iOS and Android)\, three smartphones\, a VR box\, and an impres sive collection of six dozen books.\n\n## Pre-Talk Reading List\n[Some blo gs\,concepts to go through before the talk](https://medium.com/@pacificleo /on-conversational-interfaces-notes-on-making-humanish-bots-aaac0bb901a2)\ n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T071216Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/intra-action-rethinkin g-ui-in-the-age-of-ai-Hqr1SXVWgHF59Yf93LKpxw BEGIN:VALARM ACTION:display DESCRIPTION:Intra action: rethinking UI in the age of AI in Agentic AI tra ck (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T050500Z DTEND:20250719T051500Z DTSTAMP:20260421T080241Z UID:session/WDzgg9ypPUcni8cvYpbWMj@hasgeek.com SEQUENCE:2 CREATED:20250704T153901Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164108Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Data Engineering and Infrastructure track (Semin ar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:A pragmatic guide to robust data quality checks DTSTART:20250719T051500Z DTEND:20250719T055000Z DTSTAMP:20260421T080241Z UID:session/DShS3uyEzSkZcTNBvrqF6n@hasgeek.com SEQUENCE:10 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Data & ML Infrastructure track CREATED:20250704T152130Z DESCRIPTION:Ignoring data quality introduces significant risks such as fla wed insights and poor business outcomes. This 30-minute talk moves beyond simplistic validation\, offering a journey through a multi-tiered data qua lity assurance approach. We’ll start with foundational checks: schema va lidation\, data volume monitoring\, and defining value ranges (high/low th resholds) for immediate outlier detection. These establish baseline reliab ility.\n\nNext\, the presentation explores advanced techniques for ensurin g high-quality data. It covers anomaly-based checks for unusual pattern id entification. We’ll address “gradual drift”—subtle data distributi on changes impacting models or analytics. The talk also covers inter-datas et consistency (e.g. different datasets matching the expectations that we ’d have in terms of overlap ) and strategies for data uniqueness and ded uplication.\n\nThis talk is designed to equip attendees with a clear under standing of how to layer these different checks to create a comprehensive and resilient data quality framework. The presentation will cover practica l implementation considerations within modern data stacks and showcase ill ustrative examples using popular open-source data quality frameworks. \n\n #### Key Takeaways:\n\n1. Attendees will gain a clear framework for implem enting a multi-layered data quality strategy\, progressing from basic vali dation to advanced anomaly and drift detection.\n2. Participants will unde rstand how to identify and address common data quality pitfalls\, includin g inter-data inconsistencies and data duplication\, leading to more trustw orthy data assets\, with insights into practical tooling.\n\nIf you've str uggled with poor quality data leading to data cascades\, this talk is for you.\n\n#### Speaker Bio\n\nAnay Nayak is a Solution Consultant at Sahaj S oftware with over 19 years of experience driving innovation and success in the design and delivery of large-scale enterprise projects across diverse domains. Over the last 6+ years\, he has been actively working on buildin g data platforms and integrating data science models to deliver reliable a nd actionable business insights. \n\n#### Pre - Talk Reading List\n[Few bl ogs\, concepts to go through before the talk](https://docs.google.com/docu ment/d/1YxEvyk3JpL8huVSEleBrZyd2ZRizWxIs9mbNCNVXnjc/edit?tab=t.0#heading=h .5kq2oor9cxdy) GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T071353Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/a-pragmatic-guide-to-r obust-data-quality-checks-DShS3uyEzSkZcTNBvrqF6n BEGIN:VALARM ACTION:display DESCRIPTION:A pragmatic guide to robust data quality checks in Data Engine ering and Infrastructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Morning break DTSTART:20250719T052500Z DTEND:20250719T060000Z DTSTAMP:20260421T080241Z UID:session/VqsHugBb185jpW1ZkfB8At@hasgeek.com SEQUENCE:6 CREATED:20250702T163722Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164056Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Morning break in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T052500Z DTEND:20250719T053500Z DTSTAMP:20260421T080241Z UID:session/Wa9x8eCAA3YXma4LWUCrjP@hasgeek.com SEQUENCE:3 CREATED:20250704T155319Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164104Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Relighting AI images DTSTART:20250719T053500Z DTEND:20250719T061000Z DTSTAMP:20260421T080241Z UID:session/TwcmgiSGSqaEHgooM6hdQN@hasgeek.com SEQUENCE:16 CATEGORIES:Next generation architectures,Speaking at the Fifth Elephant 20 25 Annual Conference,Demo - side project\; open source project\; something I have built in my org CREATED:20250702T164059Z DESCRIPTION:\n```Abstract:```\nReal Life photography is controlled by ligh t\, a messed up lighting can ruin a high quality Camera photograph. The si tuation is similar with AI generated Images\, specially when we are doing background swap or inpainting\, where foreground is coming from a differen t light source and background is having a different one\, the Images look pasty which affects realism.\nIn real life photography we fix lighting by manipulating camera angles while shooting\, but in AI Images it's all abou t post processing\, which boils down to Photoshop based approaches. But fo r the business who are building end to end AI Image delivery platforms usi ng Background Swaps\, Inpainting\, Photoshop is an unusable step\, so it b oils down to an automated process that fixes the light ! \nIn this talk we will talk about a custom algorithm which we developed at Caimera to fix t he light of an Image without a big latency and no loss to any details !\n\ n\n```Agenda:```\n\n- ```What is Re-Lighting ? and why is there a need ?`` `\n - We will first talk about the impact of lighting in AI Images by sho wing some AI generated Images with a comparison between correct and messed up lighting.\n - Discussion on how lighting impacts specially on in-pain ted images in terms of realism and over all likeness. This will prove the impact of correct lighting in AI generated Images.\n- ```Current Approache s a comparative study - Noise is the enemy```\n - We will compare result s of Caimera developed custom relighting algorithms against current commun ity and closed source approaches like IC-Light\, LBM Lighting.\n - Demo on comparative analysis on results\n - Discussion on how diffusion based approaches adds noise to re-create lighting in an Image\, and how it impa cts and changes details inside an Image.\n- ```High Level overview of Caim era’s Relighting Algorithm:```\n - Discussion on our non-noise based re lighting algorithm that detects lighting information from the background o f any Image by using Cromptons and Bling Phong Calculation.\n - How we ar e able to detect Shadow information\, Light Direction\, Color Information and other Light features from the Image - Mathematical overview.\n- ```Imp act of Light Information maps on Relighting:```\n - We use Depth\, Normal \, Specular\, Displacement and Color Albedo maps to represent light inform ation.\n - Discussion on how this map works and how its representation ma tters.\n - Mathematical overview of how we used this map in calculation o f light features from above point.\n - Small Demo on how each of these ma ps impact.\n- ```Light Overlaying process and Final hypothesis:```\n - We follow principles of Physics\, a mixture principle of wave optics and New ton’s Corpuscular theory where Light is considered as a Beta particle. W e will talk about how using the extracted maps and features are able to cr eate a Light mask that shows how the light will be travelling across the o bject which we are trying to relight.\n - Then we will discuss how using that light mask which is coming from considering light as a beta particle will interact on the surface. Here we take principles from wave optics and using the calculated shadow values and other features we get a clear pict ure of light's behaviour of reflection\, refraction etc.\n - Final overla y (small code demonstration)\n - Comparative Analysis on results\n- ```Di scussion on how to improve this model for better representation of light f eatures and overlay - Light on training ```\n- ```Conclusion```\n- ```QnA` ``\n \n```Takeaways for audience:```\n - Understanding of how to create realistic Images and how lighting and other features impacts\n - Understa nding a new custom algorithm \, inspiration to build something new and sim ilar \n\n```Target Audience :```\n - Machine Learning Engineers\, ML Pra ctitioners\n - Businesses\, startup founders\n - Product Managers\n - Students learning about Artificial Intelligence.\n\nslides: https://docs .google.com/presentation/d/1B0gKLGCaai_x1waIRIHeO-9lXPpl2CMiBnVw8ia3yh8/ed it?usp=sharing\n\n```About the Speaker:```\n\nMLE at Caimera AI\, Former M LE at Newton School\, Dark Horse\, Shell\, WRI\, Metvy. Contributed to Goo gle Tensorflow(GSOC)\, Samsung(Prism)\, IIT Patna (Projects). Founded MBK Health tech backed by supreme ventures to apply AI for early detection of cardiac diseases and create a hyper local support network for patients us ing wearables. Holding 4 patents on medical Imaging automations using AI a lgorithms\,holding multiple research papers and Indian young Achievers aw ard winner for contributions in artificial intelligence towards nation. Sp oken at Py-Bangalore\, Belgium-Py conference\, Keras Community Day 23\, Gi rlscript India Summit\, MIT TECH X \, HPAIR(delegate) etc and multiple mee tups\, hackathons and events.\n(```Linkedin```: https://www.linkedin.com/i n/anustupmukherjee/)\n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250709T045842Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/relighting-ai-images-T wcmgiSGSqaEHgooM6hdQN BEGIN:VALARM ACTION:display DESCRIPTION:Relighting AI images in Agentic AI track (Library) in 5 minute s TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Morning break DTSTART:20250719T055000Z DTEND:20250719T063000Z DTSTAMP:20260421T080241Z UID:session/HUKzEGbmHBCxAMjaXBkcAn@hasgeek.com SEQUENCE:2 CREATED:20250704T153559Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250704T154048Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Morning break in Data Engineering and Infrastructure track (Se minar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Hard won lessons from building real-time voice AI applications DTSTART:20250719T060000Z DTEND:20250719T063500Z DTSTAMP:20260421T080241Z UID:session/4F4zXpnWRcWTGDPNzdbpvR@hasgeek.com SEQUENCE:18 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Applied AI Engineering & Agentic AI track CREATED:20250702T163958Z DESCRIPTION:Realtime Voice AI has disrupted the conversational voicebot la ndscape and made it trivial to create exciting demos. Unfortunately\, depl oying useful voice AI in production is an entirely different story. In thi s talk\, we will present our findings from working with state of the art v oice AI agents.\n\nHow do we balance latency\, costs and scaling while ens uring the voicebot continues to be *helpful*? What are some common tradeof fs that teams will face? How do we manage complexity and orchestration of various actions that AI can take? It turns out the same kind of engineerin g maturity that defines the strongest tech teams also helps navigate the u ncharted territory in the volatile field of voice AI.\n\n## Takeaway\nGive a sense of the current challenges that product\, engineering and business teams will face when deploying cutting-edge realtime voice agents\, with pointers on how to manage these challenges.\n\n## Intended audience\nThis talk will benefit engineering\, product\, design and business teams buildi ng on top of realtime voice AI that are looking to go beyond the tech demo phase.\n\n## Bio\nAtharva is an AI aficionado. He has been tinkering with AI of all modalities to figure out how to build products that deliver on the hype in production.\n\nWhen he isn't debugging like a stoic\, he usual ly publishes strange essays online\, creates grimy digital art and produce s questionable beats with Indian truck horns.\n\nHis first exposure to ser ious software development was with the Git project\, where he rewrote part s of the submodule functionality by emailing patches to the maintainer.\n\ nWhile at nilenso\, he helped a hyperlocal delivery startup revamp their p ayout systems and solved data integration challenges at a non-profit build ing population-scale software.\n\n## Pre-talk Reading List\n[few blogs and concepts to go through before the talk](https://docs.google.com/document/ d/1lltsoW55i27mMt9_IzpFiuAJaGKclCwNhkEks5gi3O8/edit?usp=sharing)\n## Refer ences\n\nLINK: skeletal slides\, to demonstrate the flow of the talk\nhttp s://docs.google.com/presentation/d/1gX0WovWnTwFKg3-pSPNRB9_kjQJSo_N2c0G48o SkupQ/edit?usp=sharing\n\n(may be outdated\, the slides above are a better source of truth) Link to an outline on what I intend to cover:\nhttps://w ww.notion.so/nilenso-software/5thEl-talk-notes-navigating-the-challenges-o f-voice-AI-1fb0f0425dae8085b23ee0f956cca821?pvs=4\n\nLink to a blog post I wrote in January on realtime voice AI\, covers more of the PoC side of th ings:\nhttps://blog.nilenso.com/blog/2025/01/13/i-built-an-ai-prototype-th at-can-participate-in-our-internal-meetings-in-a-week/ GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T072239Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/navigating-the-challen ges-of-realtime-voice-ai-4F4zXpnWRcWTGDPNzdbpvR BEGIN:VALARM ACTION:display DESCRIPTION:Hard won lessons from building real-time voice AI applications in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Morning break DTSTART:20250719T061000Z DTEND:20250719T065000Z DTSTAMP:20260421T080241Z UID:session/MPLi9LPZDuzBA4fgYTNfQR@hasgeek.com SEQUENCE:4 CREATED:20250704T155422Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250709T045844Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Morning break in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:And yet it moves: data quality and observability | A Birds of Feat her session DTSTART:20250719T061500Z DTEND:20250719T070000Z DTSTAMP:20260421T080241Z UID:session/URQLCXQNuX62P4QqXHUxVt@hasgeek.com SEQUENCE:15 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,Birds of Feather (BOF) discussion,Data & ML Infrastructure track CREATED:20250705T074551Z DESCRIPTION:High Quality Data remains the bedrock of any kind of data anal ysis\, be it old fashioned Business Intelligence reports or the latest mac hine learning algorithms. \n\nWhat are some latest techniques for ensurin g your data is of high quality? The volume\, velocity and variety of data are ever changing\, and so also are data quality requirements. Today there is an emphasis on aspects like protecting personal information\, complyin g with geographical restrictions and being continuously aware of all the d ata assets of your organization.\n\nTake part in this Birds of Feather ses sion to learn how practitioners are approaching Data Quality in 2025. The experts on this panel will include developers solving their customers' dat a quality issues\, to products that focus extensively on this problem and much in between! You can have freewheeling interactions with this panel an d voice your thoughts on this issue by taking part in this BoF :)\n\n## Sp ecific questions within the High Quality Data BoF session\nTODO: Elaborate on this GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T163744Z LOCATION:Birds of Feather sessions (Board Room) - Bangalore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/high-quality-data-how- are-we-doing-a-birds-of-feather-session-URQLCXQNuX62P4QqXHUxVt BEGIN:VALARM ACTION:display DESCRIPTION:And yet it moves: data quality and observability | A Birds of Feather session in Birds of Feather sessions (Board Room) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Life of PII in a sea of tokens DTSTART:20250719T063000Z DTEND:20250719T070500Z DTSTAMP:20260421T080241Z UID:session/VhUCtXZWpAn6LZnigZTNqp@hasgeek.com SEQUENCE:10 CATEGORIES:Global Risk Compliance (GRC),Speaking at the Fifth Elephant 202 5 Annual Conference,30 mins talk CREATED:20250704T154032Z DESCRIPTION:**About the talk:**\nHandling Personally Identifiable Informat ion (PII) is no longer just a compliance checkbox—it’s a core responsi bility for engineers building data-intensive platforms. Whether you’re d ealing with Aadhaar\, PAN\, credit card numbers\, or biometric data\, unde rstanding how to protect sensitive information is crucial for maintaining user trust and staying on the right side of privacy regulations like GDPR and HIPAA.\n\nThis session will start by demystifying PII—what qualifies \, why it matters\, and how misuse or exposure can impact your systems and business. We’ll walk through the two primary protection mechanisms: enc ryption and tokenization\, and explain how to choose between them based on your product’s needs. \n\nWe’ll then take it further\, into the LLMs and agentic systems world. With non-deterministic behavior\, dynamic conte xt switching\, and autonomous tool use\, these systems introduce new priva cy risks that traditional PII protection strategies may not fully cover. W e’ll explore what it takes to adapt your data protection strategies when building with Generative AI and multi-agent architectures\, and how to st ay vigilant when giving AI access to sensitive datasets.\n\nIf time permit s\, we will cover a real-world case study of a global company’s architec ture will highlight how they designed for privacy and security from the gr ound up\, especially in a public cloud environment.\n\n**Takeaways from th e session:**\n- Learn a practical decision-making framework to choose betw een encryption and tokenization for protecting PII data in your systems.\n - Understand the new challenges of handling PII when building LLM-based or agentic systems\, and how to design your architecture to prevent data lea ks or misuse.\n\n**Who this is for?**\n\nThis session is ideal for softwar e engineers\, architects\, platform engineers\, MLOps professionals\, and product teams building data pipelines\, cloud platforms\, or AI-enabled sy stems—especially those working with sensitive data or exploring GenAI an d multi-agent workflows.\n\n**About Me:**\n\nI work as a Solution Consulta nt at Sahaj.ai. At Sahaj\, I am learning to build technology platforms wi th extreme programming practice and work in high-trust and accountability technology teams.\n\nI have 16 years of Software Engineering experience fo cussed on Tech Consulting and implementation and building strong technolog y teams. I have worked on Distributed systems\, and traditional monolith p latforms and built products on top of Eclipse Platform. I have diverse exp erience in Software Engineering - building tech platforms dealing with PII Data and requiring regulatory compliance\, delivery management\, customer relationship management\, and presales. \n\n**LinkedIn Profile:**\nhttps ://www.linkedin.com/in/priyadarshanpatil/\n\n**Pre-Talk Reading List:**\n[ Few blogs/concepts to go through before the talk](https://medium.com/inspi redbrilliance/building-secure-systems-with-pii-data-protection-techniques- part-i-ac9a1119cbde)\n**References:**\n\n1. Blog link - https://medium.com /inspiredbrilliance/building-secure-systems-with-pii-data-protection-techn iques-part-i-ac9a1119cbde\n2. My toastmaster speech video - https://youtu. be/V6mbhpsWSCg?si=OTneczv2c-rB6qLS\n3. PPT Link for my DevDay Hyderabad Ta lk. I would be updating the ppt to make it shorter for 30 min talk and cov er parts about Data Security in context of LLMs & Multi-agent system - htt ps://docs.google.com/presentation/d/1hseQb1VEiT-GrGgMww3NSiMCM-fczCydkTAkA -wbw74/edit?usp=sharing\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T072520Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/architecting-for-priva cy-strategies-for-pii-data-protection-VhUCtXZWpAn6LZnigZTNqp BEGIN:VALARM ACTION:display DESCRIPTION:Life of PII in a sea of tokens in Data Engineering and Infrast ructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T063500Z DTEND:20250719T064500Z DTSTAMP:20260421T080241Z UID:session/5UP3Ne91Fk3ib4a6AXZdZz@hasgeek.com SEQUENCE:6 CREATED:20250702T163817Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250704T150722Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sponsored talk | Rovo Dev CLI: coding meets copilots in terminal DTSTART:20250719T064500Z DTEND:20250719T072000Z DTSTAMP:20260421T080241Z UID:session/9nSmX9gWvrBkpjVFw6D6b2@hasgeek.com SEQUENCE:16 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Applied AI Engineering & Agentic AI track CREATED:20250702T165430Z DESCRIPTION:In the field of software development\, the incorporation of AI into coding is transforming the way developers engage with their codebase s. Atlassian's Rovo Dev CLI is leading this charge\, providing a cutting-e dge AI-powered command line tool that boosts productivity by automatically generating code. It seamlessly integrates with the Atlassian software dev elopment ecosystem\, delivering enhanced context awareness compared to tra ditional models.\n\n## What is Rovo Dev CLI\n* A command line tool that ge nerates code using advanced AI models\, code intelligence\, and deep integ ration with local codebases and Atlassian tools.\n* Developers use natural language prompts in the terminal (e.g.\, “Generate a Python function fo r X” or “Add unit tests for this file”).\n* Supports both interactiv e and non-interactive modes for a conversational experience in generating and refactoring code\, and integrating with Atlassian tools like Jira\, Co nfluence\, and Bitbucket.\n\n## Anatomy of Rovo Dev CLI\n* Understanding C odebase: Integrates with repositories\, using local memory to understand c odebase structure\, dependencies\, and coding patterns.\n* Context Awarene ss: Combines information from code and related Jira issues to generate con textually appropriate code. Integrates with MCP servers for better context awareness (e.g.\, "what tasks are assigned to me?"\, "pick task 3 and gen erate code plan"\, "mark task as completed").\n* Plan Generation: Outlines necessary files and changes before generating actual code.\n* Code Genera tion & PR: Facilitates code generation and pull requests.\n\n## Quality\n* SWE Benchmark: An offline evaluation framework developed using publicly a vailable GitHub repositories.\n* Statsig as a comprehensive multi-layer ex perimentation framework\, enabling the execution of numerous experiments s imultaneously while minimizing costs.\n* Impact: 41.98% resolve rate acros s 2294 tasks.\n\n## Enhancing Coding Agents\n* **Data:** Off-the-shelf LLM s are typically trained using publicly available datasets\, such as those measured in SWE bench. To enhance the performance of LLMs\, it is essentia l to focus on curating valuable datasets that are not accessible in the pu blic domain.\n* **Tools:** Focuses on enhancing the agent-computer interfa ce and acquiring essential information that enables agents to execute thei r tasks effectively\, such as retrieving context from Jira and generating documents in Confluence.\n* **Context:** This is primarily a search-relate d challenge stemming from the limitations of the context window in large l anguage models (LLMs). Evaluation of solutions involve offloading search c ontext to external systems to enhance performance.\n* **Behavior:** This e ntails crafting precise prompts that help clarify requirements\, facilitat e validation\, and enable self-correction.\n\n## Impact\n* New code genera tion is 80% faster.\n* Refactoring to a new domain model is approximately 50% faster.\n\n## Future Direction\nEvolution from Co-Pilot to Auto-Pilot to Pilot.\n\n## About the presenter\nBala Nathan\, Senior Principal Engine er\, Atlassian GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T141427Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/rovo-dev-cli-advancing -ai-powered-development-with-language-models-and-contextual-memory-9nSmX9g WvrBkpjVFw6D6b2 BEGIN:VALARM ACTION:display DESCRIPTION:Sponsored talk | Rovo Dev CLI: coding meets copilots in termin al in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Democratizing large model training on smaller GPUs with FSDP DTSTART:20250719T065000Z DTEND:20250719T072500Z DTSTAMP:20260421T080241Z UID:session/Tk8d7xr5sndweoFm2Zhkcw@hasgeek.com SEQUENCE:26 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Applied AI Engineering & Agentic AI track CREATED:20250702T164216Z DESCRIPTION:## Abstract\nTraining large deep learning models like LLMs and vision transformers has traditionally required high-end GPUs with large m emory\, making them inaccessible to many. This talk explores how Fully Sha rded Data Parallel (FSDP) in PyTorch can help overcome this barrier by ena bling large model training and fine-tuning on smaller GPUs (8–16GB)\, us ing commodity hardware or affordable cloud credits.\n\nWe’ll walk throug h practical experiments showcasing what’s feasible on constrained setups using FSDP. The session will cover configuration techniques such as mixed precision\, CPU offload\, and activation checkpointing\, while analyzing trade-offs with inter-GPU communication overhead.\n\nWe'll also explore ho w FSDP pairs with LoRA and QLoRA for memory-efficient fine-tuning.\n\n### Introduction & Motivation\n1. The resource challenge in training large mod els (LLMs\, ViTs)\n2. Why this talk: democratizing access to large-model t raining\n3. What is FSDP?\n### Fundamentals of Fully Sharded Data Parallel (FSDP)\n1. What gets sharded: parameters\, gradients\, optimizer states\n 2. What gets parallelized: Data\n3. What Becomes Feasible on Small GPUs?\n ### Case studies (planned/early results)\n1. Configuring FSDP\n2. Paramete r wrapping strategies and CPU Offloading\n3. PyTorch utilities and best pr actices\n### Communication Overhead: The Hidden Cost\n1. Measuring communi cation time as GPU count increases\n2. Where FSDP starts to hurt and how t o mitigate it\n3. Trade-offs in memory\, speed\, setup complexity\n### FSD P + PEFT (LoRA and QLoRA)\n1. How parameter-efficient fine-tuning compleme nts FSDP\n2. Use cases where this pairing is especially powerful\n3. Discu ssion of memory footprint vs training stability\n### Conclusion and Future Directions\n1. Summary of what’s possible with FSDP today\n\n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250709T081404Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/scaling-down-to-scale- up-fsdp-for-training-large-models-on-small-gpus-Tk8d7xr5sndweoFm2Zhkcw BEGIN:VALARM ACTION:display DESCRIPTION:Democratizing large model training on smaller GPUs with FSDP i n Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch break DTSTART:20250719T070000Z DTEND:20250719T080000Z DTSTAMP:20260421T080241Z UID:session/KGFwdERqa3F4V1r7aag72@hasgeek.com SEQUENCE:8 CREATED:20250705T074515Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T163757Z LOCATION:Birds of Feather sessions (Board Room) - Bangalore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Lunch break in Birds of Feather sessions (Board Room) in 5 min utes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T070500Z DTEND:20250719T071500Z DTSTAMP:20260421T080241Z UID:session/QVrSzZT43iAL18dQ4kx9CQ@hasgeek.com SEQUENCE:5 CREATED:20250704T154129Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250710T051624Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Data Engineering and Infrastructure track (Semin ar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:A walk in the clouds: how Uber runs multi-cloud hybrid batch jobs DTSTART:20250719T071500Z DTEND:20250719T075000Z DTSTAMP:20260421T080241Z UID:session/5MSLDvadf5N7pHMjaGxJZJ@hasgeek.com SEQUENCE:16 CATEGORIES:Big data in enterprises,Speaking at the Fifth Elephant 2025 Ann ual Conference,30 mins talk CREATED:20250704T152220Z DESCRIPTION:### Summary\n\nMigrating data and compute from on-premises to public cloud is a complex undertaking. This complexity is significantly in creased when faced with the scale of hundreds of Petabytes of data and hal f a million workloads necessary to power business intelligence and maintai n a competitive advantage. The migration of data workloads and storage to the cloud has been a multi-year initiative at Uber\, during which we are o perating in a hybrid environment between the cloud and on-premises.\n\nThe Uber batch data platform is used by thousands of engineers\, analysts as well as city operation teams across the globe to power batch and real time data processing. It uses both open source technologies such as Presto\, A pache Spark\, Pinot\, Flink and Kafka while also maintaining its customise d solutions\, for instance we have a workflow orchestrator similar to Apac he Airflow\, experimentation notebooks similar Jupyter.\n\nThis talk will discuss the key challenges we faced in migration of our batch data stack t o the cloud and cover the tooling we built to orchestrate such a large sca le migration. Even the smallest of issues in data correctness can have cat astrophic business impact\, so we will also highlight how we have helped g uarantee data correctness before and after migration.\n\nWe will also talk about some of the tradeoffs we made\, such as when to replicate data to e nsure high availability of data including and when to read data across the network from a single primary source.\n\n### Migration tooling\n\nEven be fore moving any data or compute resources\, we worked on the right tooling to identify candidate workloads and datasets incrementally while ensuring minimum disruption to users. Considering the scale and blast radius of an y issues in the migration\, we built robust automation to detect issues du ring\, or just after\, any workload is migrated\, and perform automated ro llbacks. Throughout the migration\, we used a combination of replication a cross cloud and on-prem\, and remote data access to ensure data availabili ty for consumers.\n\nWe also built abstraction layers that would work as i ntelligent proxies for any storage/compute client calls to route them to o n-premise or cloud\, depending on where the data is available and ready fo r consumption. This makes any data movement performed by platform teams to the cloud transparent to users.\n\n### Data access and integrity challeng es in a hybrid environment\n\nTo decouple storage and compute migration\, we enabled incremental replication of data across on-prem and cloud to all ow consumer workloads to run in either environment. This brought in proble ms of eventual consistency and potential for data corruption due to confli cting writes made by replication and scheduled pipelines writers (running on Uber’s custom workflow orchestrator built on Apache Airflow). We will talk in detail about how we tackled such challenges using a centralized o bservability service that tracks the availability and consistency of data across all its copies.\n\nMind map (draft slides: https://docs.google.com/ presentation/d/1wZA_N60_Gt1vTkKPYhcoNJ4pNM9WdoywTuVBHmLLiZ0/edit?slide=id. g361fb61cc07_2_100#slide=id.g361fb61cc07_2_100)\n\n### Outline:\n\n### Int roduction and Architectural Overview:\n\nWe will briefly cover the batch d ata stack on-premise\, comparing it with the setup during and post the mig ration to the cloud\n\n### Key Challenges & Constraints\n\nThe complexity involved in the migration\, including scale\, complex data usage patterns\ , constraints in data replication and availability of duplicated compute a nd storage resources.\n\n### High level migration strategy:\n\nThe process of selection of migration candidates and the tooling and automation used to monitor migrations and the guardrails in place to roll back a migration operation in case there is unexpected behaviour of data workflows\n\n### Central migration orchestration service:\n\nThe role and responsibilities of the service acting as the SOT and has been at the heart of migration\, covering all the integration points needed to make this service have near real time data on consistency and availability of data produced in the mid st of an ongoing migration.\n\n### Preventing data corruption and processi ng incomplete data:\n\nEventual consistency of source data can lead to con sumers reading data before it is completely replicated from it primary reg ion. This can have a cascading impact as it could lead to empty or incompl ete data being processed and flow down to the business intelligence system s\, which could end up taking incorrect decisions. We will cover how we ar e re-scheduling such data reads in data workflows.\n\n### Tradeoffs and co nclusions:\n\nFinally\, we will cover the key tradeoffs we made\, includin g:\n\n1. Data replication vs remote data access\n2. Need for data comple teness with tolerance for delay vs SLA requirements with tolerance for eve ntual consistency\n3. Speed vs Safety of migration\n\nand conclude with s ome key learnings throughout the two-year migration journey so far.\n\n### Pre - Talk Reading List\n[Few blogs/concepts to go through before the tal k](https://docs.google.com/document/d/1by1nPlNT4XWgkeril5mUD2qVe0eVBr5M5sU zSUjw_MM/edit?tab=t.0#heading=h.5kq2oor9cxdy)\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T072641Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/operating-ubers-batch- data-platform-in-a-hybrid-cloud-environment-5MSLDvadf5N7pHMjaGxJZJ BEGIN:VALARM ACTION:display DESCRIPTION:A walk in the clouds: how Uber runs multi-cloud hybrid batch j obs in Data Engineering and Infrastructure track (Seminar Hall) in 5 minut es TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch break DTSTART:20250719T072000Z DTEND:20250719T082000Z DTSTAMP:20260421T080241Z UID:session/Jkr1LS7F6M49ap8PUi9qwb@hasgeek.com SEQUENCE:7 CREATED:20250702T163904Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164041Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Lunch break in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch break DTSTART:20250719T072500Z DTEND:20250719T085000Z DTSTAMP:20260421T080241Z UID:session/A3eeUfmnUspFPX4R7rT3No@hasgeek.com SEQUENCE:9 CREATED:20250704T155859Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T024723Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Lunch break in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch break DTSTART:20250719T075000Z DTEND:20250719T085000Z DTSTAMP:20260421T080241Z UID:session/RvNaD8xHEJch6DHtmUTBay@hasgeek.com SEQUENCE:4 CREATED:20250704T154201Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250710T051643Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Lunch break in Data Engineering and Infrastructure track (Semi nar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Software development in the age of AI | Birds of Feather session DTSTART:20250719T080000Z DTEND:20250719T090000Z DTSTAMP:20260421T080241Z UID:session/AGUFfgMYQrLb3E7mzCVGPy@hasgeek.com SEQUENCE:30 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,Birds of Feather (BOF) discussion,Applied AI Engineering & Agentic AI track CREATED:20250703T030339Z DESCRIPTION:Generative AI is transforming the software development workflo w. What are some current techniques\, tools and processes that harness the power of AI to great effect? And how to avoid subtle traps and pitfalls? This Birds of Feather session should hopefully have you amazed and provoke d by what the participants will say - and show - about this topic!\n\n## P otential topics in this BoF session\n1. Some developers showing workflows and patterns that are effective for them\, and discussions around it\n2. R elevant tools and services in this space. Some cost-benefit analysis\n3. H ow collaboration with other roles is also impacted? Designer -> Developer -> Testing -> Ops etc.\n4. Trendspotting of how development workflows are likely to change in next few years\n \n## Participants\nTDB GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T163806Z LOCATION:Birds of Feather sessions (Board Room) - Bangalore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/software-development-i n-the-age-of-ai-a-birds-of-feather-session-AGUFfgMYQrLb3E7mzCVGPy BEGIN:VALARM ACTION:display DESCRIPTION:Software development in the age of AI | Birds of Feather sessi on in Birds of Feather sessions (Board Room) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sponsored talk | AI innovation at the core of Google search DTSTART:20250719T082000Z DTEND:20250719T085500Z DTSTAMP:20260421T080241Z UID:session/4DkAzQ9Hs4ZUPXHYzdc4q4@hasgeek.com SEQUENCE:10 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Applied AI Engineering & Agentic AI track CREATED:20250702T163610Z DESCRIPTION:## Description of the problem statement or the context\nFor ye ars\, Google has facilitated users' information journeys with its mission to organize the world's information and make it universally accessible. T oday that mission is more relevant than ever as users are embarking on inc reasingly complex information journeys. AI is the key to unlocking unprece dented depth and utility within these journeys\, moving beyond simple answ ers to fostering genuine understanding and discovery. Let’s dive into ho w Google is both innovating and leveraging AI advancements to push the bou ndaries of search\, transforming it into an intelligent partner for billio ns of users worldwide.\n\n## Key takeaways for the audience\n1. Learn abou t state-of-the-art AI techniques and applications directly from an industr y leader\n2. Understand the capabilities\, limitations\, and ethical consi derations of powerful AI systems\n3. Beyond the buzz\, get insights into w hat it takes to build AI solutions at Google Scale\n\n## Audience for this talk \nAI Practitioners\, Engineers\, Technical and Product Leads - Learn directly from Google about their state-of-the-art AI advancements. Unders tand the practical capabilities\, crucial limitations\, and ethical dimens ions of powerful AI. Plus\, gain unparalleled insights into the strategies and challenges of building and scaling AI solutions like Google does. Ele vate your understanding and approach to AI development.\n\n## Bio\nShruti is an AI lead at Google leading initiatives to expand Search's frontiers w ith multimodal AI. She has a decade worth of experience on solving real-wo rld problems in the NLP and Vision domains across different Google product s. GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T141707Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/breaking-barriers-how- ai-is-pushing-the-limits-of-google-search-4DkAzQ9Hs4ZUPXHYzdc4q4 BEGIN:VALARM ACTION:display DESCRIPTION:Sponsored talk | AI innovation at the core of Google search in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Bridge across forever: unifying enterprise metadata DTSTART:20250719T085000Z DTEND:20250719T092500Z DTSTAMP:20260421T080241Z UID:session/G3R6dzjcgmq27pxCx2wa9t@hasgeek.com SEQUENCE:12 CATEGORIES:Managing meta data,Speaking at the Fifth Elephant 2025 Annual C onference,30 mins talk CREATED:20250704T160812Z DESCRIPTION:The rise of AI has transformed data into a strategic asset\, y et traditional metadata systems are failing to keep pace. Passive catalogs \, built for simpler times\, crumble under the demands of scalable\, trust ed AI. We're entering an era where metadata must be active\, trusted\, and governed – becoming the central nervous system for your data ecosystem. \n\nThis talk introduces the Metadata Lakehouse (MLH)\, a revolutionary ar chitectural paradigm we're pioneering and actively implementing at Atlan. Tailored for data architects\, engineers\, and leaders\, this session appl ies the proven principles of the data lakehouse to metadata\, demonstratin g how the MLH unifies all enterprise metadata on open\, scalable cloud sto rage. We'll delve into how technologies like Apache Iceberg provide ACID t ransactions\, schema evolution\, and time travel for metadata\, moving bey ond simple lookup to enable:\n\n- Architecting the Future: Understand the fundamental limitations of traditional metadata systems and the core archi tectural principles of a Metadata Lakehouse (MLH).\n- Empowering AI Govern ance: Learn how MLH enables advanced metadata analytics\, verifiable linea ge\, bias management\, and risk tracking for AI/ML assets.\n- Unifying You r Data World: Discover strategies for true silo unification\, consolidatin g metadata for end-to-end lineage and consistent policy enforcement.\n- Dr iving Automation: Explore how MLH serves as the foundation for metadata-dr iven AI\, intelligent discovery\, and automated governance.\n\nJoin us for a deep dive into the practical architectural blueprints and real-world en gineering challenges of building this next-generation metadata infrastruct ure. Our experience at Atlan\, building sophisticated data governance plat forms\, provides unique insights into the practicalities and opportunities of this vision. \n\nWe'll candidly discuss the implementation realities a nd strategic implications for establishing the essential control plane for data and AI in the enterprise. Understand why the Metadata Lakehouse is n ot just an evolution\, but a necessity for unlocking your organization's f ull AI potential. GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T025837Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/metadata-lakehouse-bui lding-the-active-nervous-system-for-enterprise-ready-ai-control-plane-G3R6 dzjcgmq27pxCx2wa9t BEGIN:VALARM ACTION:display DESCRIPTION:Bridge across forever: unifying enterprise metadata in Data En gineering and Infrastructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sponsored session | The great pivot: from satellites to scale DTSTART:20250719T085000Z DTEND:20250719T093500Z DTSTAMP:20260421T080241Z UID:session/WpzAdoH12iX8uPVuNM5t7s@hasgeek.com SEQUENCE:4 CREATED:20250709T045506Z DESCRIPTION:### 🎯 Opening flow\n* **Setup:** “Meet someone who actual ly WAS a rocket scientist...”\n* **Hook:** 24 years tracking satellites → carts → custom AI\n* **Big Question:** Why leave ISRO?\n* **Key stor y:** Moment he decided to leave ISRO\n\n---\n\n## 🔷 Block 1: Career Tra nsition (8–12 mins)\n\n### 💡 Key Themes\n* Transferable skills: mathe matical modeling → ML\n* Mindset shift: deterministic → probabilistic\ n* Culture shift: ISRO vs Flipkart\n* Personal growth: stakeholder mgmt\, business language\n\n### 🧩 Audience Takeaway\n\n* Cross-disciplinary tr ansitions are possible\n* Precision meets business ambiguity\n\n---\n\n## 🔷 Block 2: Problem-Solving Methodology (10–15 mins)\n### 🧱 Key Poi nts\n* First principles over paper-copying\n* Visiting the field to unders tand real-world challenges\n* Framing vague problems mathematically\n\n### 🧪 Case Studies\n* Indian address fraud detection\n* Delivery predictio n and promise modeling\n* ISRO: orbital calculation precision\n\n### 🧠 Audience Takeaway\n* First principles > replication\n* Understanding groun d > fancy algorithm\n\n---\n\n## 🔷 Block 3: AI Implementation Challenge s (12–18 mins)\n\n### 🏗️ Topics\n* Multi-domain AI at Sahaj\n* Cust om vs generic AI solutions\n* Reasons AI fails: stakeholder buy-in\, laten cy\, scale\n* Practical deployment lessons: adoption > algorithm\n* Pragma tism: simple solutions often win\n\n### 🧠 Audience Takeaway\n* AI succe ss = deployment\, not sophistication\n\n---\n\n## 🔷 Block 4: Career Gui dance (5–8 mins)\n\n### 🔮 Future Skills\n* Beyond math: consulting\, domain understanding\n* Industry-ready researchers\n* Product-minded data scientists\n* Becoming translators between tech and business\n\n### 💬 C ontrarian Advice\n* Not everyone should code\n* Specialize vs generalize\n * Startup vs enterprise tracks\n\n### 🧠 Audience Takeaway\n* Diverse ba ckgrounds & problem framing are key\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T145649Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/the-great-pivot-from-s atellites-to-scale-WpzAdoH12iX8uPVuNM5t7s BEGIN:VALARM ACTION:display DESCRIPTION:Sponsored session | The great pivot: from satellites to scale in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T085500Z DTEND:20250719T090500Z DTSTAMP:20260421T080241Z UID:session/BrQb1ZUjnj9mptzLEXBNTZ@hasgeek.com SEQUENCE:5 CREATED:20250702T164009Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250704T150731Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T090000Z DTEND:20250719T090500Z DTSTAMP:20260421T080241Z UID:session/NRF2UGbceYma8pSB2jz1Jc@hasgeek.com SEQUENCE:5 CREATED:20250716T070738Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164031Z LOCATION:Birds of Feather sessions (Board Room) - Bangalore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Birds of Feather sessions (Board Room) in 5 minu tes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Speaking dev to dev: crafting talks that stick DTSTART:20250719T090500Z DTEND:20250719T095000Z DTSTAMP:20260421T080241Z UID:session/6d3oJVDovUCrr3pbk6gGix@hasgeek.com SEQUENCE:6 CREATED:20250716T070721Z DESCRIPTION:You may have a great idea you want to talk about. But you are struggling to create a presentation around it. \n\nIn this session\, I w ill present guidelines to help you create an impactful talk. It will give you the tools you need to submit your talk to the next [Fifthel](https:// hasgeek.com/fifthelephant) or [Rootconf](https://hasgeek.com/rootconf) con ference.\n\n# Outline\n\n1. What makes a good talk ?\n- not a survey\n- no t publicly available\n- something non-trivial\n- experience report or in-d epth analysis of some tool\, algorithm\, Data Science.\n\n2. Understanding the human mind\n- how much can it remember\n- what does it remember\n\n3. Add quiz - involve the audience\n\n4. Not too many colors\n- Use consiste nt color scheme \n- Change the font size\n\n5. How much text\n- Pictures \ n\n6. Aristotle\n- A good script\n\n7. Agenda : let the audience know what to expect\n\n8. Takeaways\n- Start with a shocking or intriguing fact\n- Divide into Sections\n- Reinforcing the message - Summarize again\n- Backu p slides\n\n9. What goes wrong\n- Too much detail - let me explain everyth ing\n- Not enough detail - let me explain big picture GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T163849Z LOCATION:Birds of Feather sessions (Board Room) - Bangalore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/how-to-walk-the-talk-6 d3oJVDovUCrr3pbk6gGix BEGIN:VALARM ACTION:display DESCRIPTION:Speaking dev to dev: crafting talks that stick in Birds of Fea ther sessions (Board Room) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Enhancing nuclear safety with vision-based AI DTSTART:20250719T090500Z DTEND:20250719T094000Z DTSTAMP:20260421T080241Z UID:session/GsgBGwv9cRqTZ5ky8xyHNq@hasgeek.com SEQUENCE:10 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Visual AI track CREATED:20250702T163527Z DESCRIPTION:Hydrogen explosions are an inherent risk in nuclear reactors d ue to the unavoidable generation of hydrogen during its operation. From Ch ernobyl to Fukushima\, several major nuclear accidents have been significa ntly worsened by hydrogen explosion. In sodium-cooled fast breeder reactor s\, due to the use of sodium as a coolant\, hydrogen production is prevale nt. The disposal of used radioactive sodium involves spray injection\, whi ch can unintentionally lead to hydrogen build up and explosions if not pro perly controlled.\n\nThis talk presents how Visual AI techniques are used to monitor and mitigate such risks. Through real-time video analysis\, dee p learning models detect early signs of hydrogen accumulation and flame an omalies. These insights are then used to trigger automated control systems based on reinforcement learning techniques to optimize the disposal opera tion\, enabling intelligent decision-making in dynamic environments. The s ession will cover data generation\, model development\, and how AI solutio ns are deployed in safety critical nuclear applications.\n\n**Outline of T alk**\n- Introduction to Fast Breeder Reactors (3 mins)\n - *Overview of three stage nuclear program*\n - *Role of fast breeder reactors and its c hallenges*\n- Need for Sodium Disposal (2 mins)\n - *Why disposal of used radioactive sodium is essential*\n - *Spray injection method and its ris ks*\n- Sodium Fire Detection (8 mins)\n - *Data generation: Controlled ex periments simulating sodium fires using experiments*\n - *Data annotation : Building labelled datasets from raw footage*\n - *Modeling: Training De ep learning models for real time fire detection*\n- Video captioning of hy drogen deflagration event (6 mins)\n - *Real time video analysis*\n - *R NN based models to caption the events*\n- Reinforcement Learning for auton omous control (5 mins)\n - *Reinforcement learning to optimize the spray injection process*\n - *Intelligent decision-making to prevent hydrogen a ccumulation*\n- Future Work (1 min)\n - *Deployment on edge device*\n \n* *Key Takeaways**\n• Role of AI in Nuclear Safety\n• Experimental data generation in industrial environment\n• End to End pipeline of Vision ba sed automation system\n\n**Speaker Info**\nMuthukumar Ganesan is currently working as a Scientist at the Atomic Research Centre\, Government of Indi a. He brings over 11 years of professional experience in AI application de velopment across the automotive and nuclear industries. His work focuses o n enhancing safety and automation in sodium-cooled fast breeder reactors u sing AI-driven solutions\, including hydrogen explosion detection\, sodium leak identification\, predictive maintenance\, and autonomous fire mitiga tion systems. He has published 3 research papers in the field of AI and is currently working on 2 more. An experienced technical speaker\, he has de livered talks on AI applications in industrial systems and actively contri butes to the GitHub and MathWorks communities\, where he shares innovative tools and solutions for real-world engineering challenges.\n[**LinkedIn** ](https://www.linkedin.com/in/muthukumar-ganesan-b6b85461/)\n[**Git**](htt ps://github.com/muthuganeshece)\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T072735Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/ai-for-nuclear-safety- vision-based-reinforcement-learning-for-the-prevention-of-safety-critical- event-in-reactor-facilities-GsgBGwv9cRqTZ5ky8xyHNq BEGIN:VALARM ACTION:display DESCRIPTION:Enhancing nuclear safety with vision-based AI in Agentic AI tr ack (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T092500Z DTEND:20250719T093500Z DTSTAMP:20260421T080241Z UID:session/Uo8Cr2Zysfh1ZanYypJ1k@hasgeek.com SEQUENCE:4 CREATED:20250704T154259Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164025Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Data Engineering and Infrastructure track (Semin ar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T093500Z DTEND:20250719T094500Z DTSTAMP:20260421T080241Z UID:session/TpPtDSGkgQ88ifvGAStA4n@hasgeek.com SEQUENCE:5 CREATED:20250705T074003Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T164017Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:From data to dialogue: making databases conversational and intelli gent DTSTART:20250719T093500Z DTEND:20250719T101000Z DTSTAMP:20260421T080241Z UID:session/L4ioGPTEVQ24KSRYoKyc5G@hasgeek.com SEQUENCE:9 CATEGORIES:LLM-Ops,Speaking at the Fifth Elephant 2025 Annual Conference,3 0 mins talk CREATED:20250704T160434Z DESCRIPTION:## Overview\n\nAs organizations race to integrate large langua ge models into their products and workflows\, a new requirement is emergin g: the need to host private LLMs in a scalable\, secure\, and operationall y manageable way.\n\nThis talk presents a practical\, cloud-agnostic archi tecture for hosting private LLMs with strong security isolation and effici ent AI operations at scale.\n\nWe'll explore how to enforce isolation\, es tablish secure private network boundaries\, and build a hardened control p lane to manage LLM lifecycle and infrastructure state.\n\nCentral to this architecture is a model-agnostic access layer or a gateway\, which decoupl es downstream systems from specific model APIs. It provides a consistent i nterface across model types and versions\, while enabling operational feat ures like request authentication\, batching\, standard and semantic cachin g\, and routing.\n\nIn addition to architecture\, we’ll explore the oper ational challenges of managing private LLMs in production\, such as GPU re source scaling\, long-tail latency under concurrent load\, scaling under u npredictable traffic\, and cost optimization techniques. \n\n\n---\n\n\n## Takeaways\n\nAttendees will learn:\n\n\n\n* How to design a secure isolat ion layer for private LLMs using cloud-native constructs.\n* How to implem ent private\, low-latency access using cloud-native networking primitives. \n* The role of a model-agnostic AI Gateway (Access Layer) in:\n * Unif ying access across different LLM backends\n * Managing API key auth and RBAC\n * Implementing standard\, semantic and conversational caching\n * Aggregating requests for efficient batching\n* Operational strategie s for:\n * Orchestration and Upgrades\n * Reducing long-tail latency \n * Controlling cost under bursty traffic\n * Autoscaling Strategie s\n * Performance and Cost Tradeoffs\n\n\n\n---\n\n\n## Audience\n\nThi s session is designed for:\n\n\n\n* Platform Engineers building secure AI infrastructure\n* MLOps / DevOps Engineers managing the deployment and sca ling of LLMs\n* Cloud Infra and SRE Teams responsible for performance\, av ailability\, and cost control\n* AI Engineers deploying private models in enterprise\, internal\, or regulated settings\n* Anyone designing or runni ng LLM infrastructure beyond prototypes\n\n---\n\n## The Challenge: Why Mu lti-Tenant LLM Serving is Hard (Problem Statement)\n\n\n\n* Deploying LLMs effectively for multiple customers (tenants) goes beyond simple model hos ting.\n* Key challenges: Ensuring scalability\, robust security\, tenant d ata isolation\, cost management\, and integrating value-added features.\n* Existing solutions often lack integrated\, enterprise-grade capabilities\ , forcing organizations to build complex frameworks themselves for:\n * Performance/Cost Optimization (Caching\, Batching)\n * Model Agnostici sm & Upgrades\n * Operational Needs (Auth\, RBAC\, Monitoring\, Secure Networking)\n\n\n## Our Solution: A Layered AWS Architecture (Overview)\n\ n\n\n* Presenting a robust\, multi-tenant LLM platform architecture built on AWS.\n* Designed for scalability\, security\, cost-efficiency\, and eas e of use for tenants.\n\n\n## Core Principle: Secure Tenant Isolation\n\n\ n\n* **Strategy:** AWS Account-per-Customer.\n* **Benefits:**\n * Stric t data separation and isolation.\n * Simplified per-tenant billing and cost tracking.\n * Enables secure\, customer-specific networking (Priva teLink).\n * Facilitates meeting compliance requirements.\n* Managed vi a a central control plane using cross-account IAM roles.\n\n\n## Key Archi tectural Components\n\n\n\n* **Model Serving Layer:**\n * Leverages opt imized toolkits (e.g.\, vLLM\, NVIDIA NIM) for standardized inference APIs and performance.\n * Model Serving Agent (on EC2): Manages model lifec ycle (deploy\, start/stop\, update)\, reports health\, collects metrics (f or CloudWatch/Prometheus)\, and routes requests.\n* **Networking & Secure Access:**\n * **Public Access:** Standard ALB + ASG + Route53 setup for stable public endpoints.\n * **Private Access (Preferred):** AWS Priva teLink for secure\, private connectivity from customer VPCs to the LLM ser vice (avoids CIDR conflicts\, simplifies security).\n\n\n## AI Gateway: Th e Value-Add Layer:\n\n\n\n* An intermediary service providing crucial feat ures before hitting the model server.\n* **Authentication & Authorization: ** API Key management\, Role-Based Access Control (RBAC).\n* **Intelligent Caching:**\n * **Standard Caching:** Key/Value store for identical pro mpts.\n * **Semantic Caching:** Vector DB lookup for similar/paraphrase d prompts.\n* **Request Batching:** Aggregates requests for improved throu ghput and cost-efficiency (especially if not native to the toolkit).\n\n\n ## Smart Scaling & Cost Management\n\n\n\n* **Challenge:** Standard metric s (CPU/Network) don't accurately reflect LLM load (GPU is the bottleneck). \n* **Solution:** GPU-Centric Auto Scaling using AWS Auto Scaling Groups ( ASGs).\n * Collect GPU Utilization (%) via nvidia-smi on instances.\n * Publish as Custom CloudWatch Metrics.\n * Configure ASG Scaling Pol icies (Target Tracking/Step Scaling) based on these custom GPU metrics.\n* **Benefits:** Accurate scaling\, better performance\, cost optimization b y avoiding over/under-provisioning.\n* **User Control:** Allow tenants to enable/disable auto-scaling and set min/max instance limits.\n\n## Benefit s & Use Cases\n### AI Functions\nWith the current set of databases\, custo mers can only access data in a structured format. If they want to leverage that data with LLMs\, they typically need to write a custom client that r etrieves data from the database and sends it to the LLM.\n\nCouchbase Serv er already supports User-Defined Functions (UDFs) via SQL++. We utilized U DFs to directly invoke LLMs and return the responses to the user.\n\nHowev er\, we encountered a challenge: UDFs do not natively support authenticati on. While we had firewalls in place for our AI functions\, relying on fire walls alone is not sufficient for robust security. To address this\, we im plemented AWS STS to generate temporary tokens\, providing an additional l ayer of secure access.\n\n### Vectorization Service\nNow that we support e mbedding models\, we wanted to provide customers with a way to vectorize t heir existing data in Couchbase Server.\n\n#### Introduction to DCP\nCouch base Server includes its own protocol\, DCP (Database Change Protocol)\, w hich streams document mutations to clients. One such client is the Eventin g Service—an existing feature in Couchbase that allows users to write cu stom JavaScript logic to handle document mutations.\n\nTo deliver a seamle ss experience without reinventing the wheel\, we chose to leverage the Eve nting Service (already a DCP consumer) to vectorize customer data efficien tly.\n\n### UDS(Unstructured Data Service)\nWe also provided customer's a way to add data to the database from PDF\, text document etc. We created o ur own service UDS that extract JSON documents from these file and then in sert them into the database.\n### Agent Catalog\nWe wanted to provide cust omers a way to query data using natural language by using Agent catalog.\n \nAgent catalog manages their queries\, they can integrate their own agent s to this and when they give a agent catalog can perform a vector search a nd find the most relevent query and the agents can then execute that query .\n\n### Pre-talk Reading List\n[Few concepts/blogs to go through before t he talk](https://docs.google.com/document/d/1YWtO6c82_q39QejCrxznHV9oFJ5I9 e28gNzSOpcicDk/edit?tab=t.0)\n## Conclusion\n* Building a successful multi -tenant LLM platform requires thoughtful architecture beyond basic deploym ent.\n* Combining AWS best practices (Account-per-Tenant\, PrivateLink) wi th custom components (Model Serving Agent\, AI Gateway) and intelligent sc aling (GPU metrics) delivers a powerful solution.\n* Empowers customers to leverage LLMs securely and efficiently without managing the underlying co mplexity. GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T072856Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/from-data-to-dialogue- making-databases-conversational-and-intelligent-L4ioGPTEVQ24KSRYoKyc5G BEGIN:VALARM ACTION:display DESCRIPTION:From data to dialogue: making databases conversational and int elligent in Data Engineering and Infrastructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T094000Z DTEND:20250719T095000Z DTSTAMP:20260421T080241Z UID:session/16oX1FERQfUtvzm7yg2gne@hasgeek.com SEQUENCE:3 CREATED:20250704T150651Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250704T150930Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Demos and Flash talks DTSTART:20250719T094500Z DTEND:20250719T103000Z DTSTAMP:20260421T080241Z UID:session/MnCKsRNTyRVPqnK777kNdi@hasgeek.com SEQUENCE:12 CREATED:20250705T074657Z DESCRIPTION:Sign up to present your work as a 5-min Flash talk or Demo!\nR egister at the nearest Helpdesk / Registration desk to grab a spot! GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T145921Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Demos and Flash talks in Agentic AI track (Library) in 5 minut es TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T095000Z DTEND:20250719T095500Z DTSTAMP:20260421T080241Z UID:session/GZbWZ6tbDCRgBz8C3vjVB8@hasgeek.com SEQUENCE:1 CREATED:20250717T163918Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T163933Z LOCATION:Birds of Feather sessions (Board Room) - Bangalore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Birds of Feather sessions (Board Room) in 5 minu tes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Make your videos speak perfectly - with advanced lip-sync AI DTSTART:20250719T095000Z DTEND:20250719T102500Z DTSTAMP:20260421T080241Z UID:session/Ct1R2K5TURoiwecjEJ789u@hasgeek.com SEQUENCE:14 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Visual AI track CREATED:20250702T165155Z DESCRIPTION:We present a production-ready lip-sync model serving millions of creators\, built on a novel Latent-GAN architecture that achieves super ior identity preservation and audio-visual alignment compared to other app roaches.\n\nOur system is trained on 10\,000+ hours of diverse audio-visua l data using custom preprocessing pipelines which include audio diarizatio n\, vocal separation\, and AV synchronization etc. We demonstrate how GAN architectures with transformer attention mechanisms and VAEs can match dif fusion model quality while offering faster inference speeds.\n\nKey techni cal contributions include:\n- Latent-GAN architecture leveraging transform er blocks and VAE improvements for high-resolution output\n- Custom loss f unctions for facial feature consistency and identity preservation across d iverse demographics\n- Scalable preprocessing pipeline handling 10TB+ of h eterogeneous audio-visual data.\n- Production inference system built in El ixir\, achieving sub-second response times at scale\n- Integration framewo rk with existing B-roll generation and AI director systems\n- Solutions fo r common lip-sync challenges: temporal coherence\, cross-identity generali zation\, and multi-speaker scenarios\n\nTarget audience: ML engineers\, AI researchers\, and developers building content creation tools\, video proc essing systems\, or scaling AI models for consumer applications.\n\nhttps: //pitch.com/v/fifth-elephant---2025-uwv9ug\n\n## Pre - Talk Reading List\n [Few blogs/concepts to go throuhg before the talk](https://docs.google.com /document/d/17oqYvM0jEVeFFokY-_j5sEljn_8gswXHVRlcPfOOvfU/edit?tab=t.0#head ing=h.5kq2oor9cxdy) GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T073014Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/revolutionise-content- creation-with-advanced-lip-sync-ai-Ct1R2K5TURoiwecjEJ789u BEGIN:VALARM ACTION:display DESCRIPTION:Make your videos speak perfectly - with advanced lip-sync AI i n Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Finding signal in a noisy landscape: keeping up with AI developmen ts | Birds of Feather session DTSTART:20250719T095500Z DTEND:20250719T104000Z DTSTAMP:20260421T080241Z UID:session/Tcp81Vfz49e5VTvTLNWQnJ@hasgeek.com SEQUENCE:17 CREATED:20250709T045818Z DESCRIPTION:Generative AI has been the fastest moving technology today. It also happens to be terribly misunderstood.\nHow can we build a clear unde rstanding of AI capabilities? How do we curate an information environment that lets us accurately assess the capabilities of the technology? What de velopments are worth paying attention to?\nJoin this session to find out!\ n\n ## Potential topics for this session:\n* How do we keep up with all th at fast-moving AI progress?\n* How do we educate ourselves to understand t his technology deeply?\n* Where can we find useful signals that will help us make responsible business and technical decisions? How do we avoid nois e and undue hype?\n* What can we do as technologists to create a good info rmation environment? How do we communicate the capabilities of AI usefully ?\n* How can we advance the field of AI with the resources we have?\n\n## What you'll take away:\n* A framework for evaluating AI capabilities and s eparating genuine breakthroughs from hype\n* Practical strategies for buil ding your own AI information diet (specific sources\, tools\, and filterin g techniques)\n* Methods for translating technical AI developments into ac tionable business and product decisions\n* A curated list of high-signal r esources and communities worth following\n\n## Who should attend:\n* Techn ical leaders and practitioners who want to keep pace with AI developments and distinguish signal from noise\n* Those who've developed effective stra tegies for staying current and want to exchange notes on how they upskill and monitor the frontier\n* Engineers\, product managers\, architects\, an d technical decision-makers who need to make informed choices about AI ado ption.\n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T163904Z LOCATION:Birds of Feather sessions (Board Room) - Bangalore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/birds-of-feather-bof-s ession-finding-signal-in-a-noisy-landscape-keeping-up-with-ai-developments -Tcp81Vfz49e5VTvTLNWQnJ BEGIN:VALARM ACTION:display DESCRIPTION:Finding signal in a noisy landscape: keeping up with AI develo pments | Birds of Feather session in Birds of Feather sessions (Board Room ) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transition DTSTART:20250719T101000Z DTEND:20250719T102000Z DTSTAMP:20260421T080241Z UID:session/97zwKc8RZEH1AMbbuCjCxB@hasgeek.com SEQUENCE:3 CREATED:20250704T155109Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250710T051652Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transition in Data Engineering and Infrastructure track (Semin ar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Zasper: a high performance JupyterLab alternative DTSTART:20250719T102000Z DTEND:20250719T105500Z DTSTAMP:20260421T080241Z UID:session/DCLSWaiYmHUz5xoT1RkFTo@hasgeek.com SEQUENCE:21 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Data & ML Infrastructure track CREATED:20250704T152149Z DESCRIPTION:# Description\nData science tools have come a long way\, and P roject Jupyter has been foundational to that progress. But what if we coul d dramatically improve their performance without abandoning the Python eco system?\n\nIn this talk\, I’ll introduce Zasper\, a high-performance IDE for Jupyter notebooks that delivers:\n\n* Up to 5× lower CPU usage\n* Up to 40× lower RAM usage\n* Lower latency and higher throughput\n* Massive concurrency support with minimal memory overhead\n\nZasper achieves this by reimplementing parts of the Jupyter server stack in Go\, while staying fully compatible with the Jupyter protocol. If you’ve ever hit performan ce bottlenecks with traditional tools\, this talk is for you.\n\n# Takeawa ys\n\nZasper is a reimagined IDE for Jupyter notebooks\, designed from the ground up with high performance and concurrency in mind. It maintains com patibility with Jupyter’s wire protocol while replacing Python-based com ponents with lean\, efficient Go implementations.\n\nThis talk will cover: \n\n* How the Jupyter server and protocol work under the hood\n* Architect ural pain points in traditional Python-based implementations\n* Where Go c an be introduced without compromising Python workflows\n* Benchmark compar isons between JupyterLab and Zasper\n* Lessons learned from building a Go- based Jupyter-compatible server\n\nBy the end of the session\, attendees w ill have a deeper understanding of the internals of Jupyter\, and how comb ining Go and Python can unlock a new class of high-performance PyData tool s—ideal for large-scale\, multi-user\, or production-grade notebook envi ronments.\n\n# Audience\nIntermediate\n\n# Bio\nPrasun loves building open -source software for scientific computing. He is the creator of Zasper\, a high-performance IDE for Jupyter Notebooks.\n\n# Pre-Talk Reading List\n( Few blogs/concepts to go through before the talk)[https://docs.google.com/ document/d/1pXxUxPqvO5NSnIwsbnQwHZoz5fsVeZRUB36qRyJtgFw/edit?tab=t.0#headi ng=h.5kq2oor9cxdy]\n# Reference\n1. [Project](https://github.com/zasper-io /zasper)\n2. [Benchmark code and Report](https://github.com/zasper-io/zasp er-benchmark)\n# Draft Slides & Demo\n[Slides](https://drive.google.com/fi le/d/1bpgwfJfd-4ohwtYrkqPWWn2DTlt6-GHI/view?usp=sharing)\n[Demo](https://w ww.youtube.com/watch?v=LvVOkYL_LzQ)\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T073136Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/building-high-performa nce-pydata-tools-by-adding-go-to-the-mix-DCLSWaiYmHUz5xoT1RkFTo BEGIN:VALARM ACTION:display DESCRIPTION:Zasper: a high performance JupyterLab alternative in Data Engi neering and Infrastructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Evening break DTSTART:20250719T102500Z DTEND:20250719T105500Z DTSTAMP:20260421T080241Z UID:session/XmiGXdWrUEEwazsCinFxsE@hasgeek.com SEQUENCE:5 CREATED:20250702T164142Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T074214Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Evening break in Agentic AI track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Evening break DTSTART:20250719T103000Z DTEND:20250719T105500Z DTSTAMP:20260421T080241Z UID:session/F1f5Lcf62YfhziAY1tPtPD@hasgeek.com SEQUENCE:11 CREATED:20250704T160046Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250711T131840Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Evening break in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Enhancing retrieval in RAG: the fused features way DTSTART:20250719T105500Z DTEND:20250719T113000Z DTSTAMP:20260421T080241Z UID:session/4MGFKwLnTjXgu8RnzEtisG@hasgeek.com SEQUENCE:12 CATEGORIES:Experimentation,Speaking at the Fifth Elephant 2025 Annual Conf erence,30 mins talk CREATED:20250709T045644Z DESCRIPTION:Retrieval-Augmented Generation (RAG) has emerged as a dominant framework for leveraging Large Language Models (LLMs) to generate respons es grounded in extensive textual corpora. Within this architecture\, the r etrieval component plays a critical role in determining the overall system accuracy by surfacing the most relevant text chunks based on semantic sim ilarity to the user query. Typically\, this similarity is computed via cos ine scores between vector embeddings of the query and document segments. T his talk will highlight the limitations of conventional retrieval methods and motivate the need for more expressive and effective embedding strategi es.\n\nWe will look into both sparse and dense embeddings\, and how each c aptures different aspects of meanings from text. The talk will focus on ho w combining these embeddings can give a more complete representation of qu eries and documents. I will explain simple yet powerful techniques to fuse multiple embeddings and show how this improves retrieval results. Through practical examples and empirical insights\, the session will demonstrate how such fusion techniques significantly outperform single-embedding basel ines in RAG pipelines.\n\nOutline of the talk\n- Introduction to embedding s as effective representation of text\n- What are sparse and dense embeddi ngs and what do they really represent \n- Ways to combine multiple embeddi ngs to form fused or composite features\n- Retrieval scores in RAG\, showc asing the effectiveness of fused features\n\nTakeaways\n- Learn about unde rlying mechanisms of text embeddings\n- Learn cool ways to make features f rom text\; simple refinements that make huge difference in RAG\n- Listen t o nuances of some original work that I have done for in-house projects\n\n If you're a Gen AI enthusiast\, going to build many many RAG systems and h ave curiosity around LLMs\, this talk is for you!\n\nSpeaker bio\nDr. Kart hika Vijayan is a Solution Consultant at Sahaj Software. She has been cond ucting research in the field of conversational AI with voice and text data for almost a decade. Her research has been published in several journals and presented at various international conferences. 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\nhttps://www.youtube.com/watch?v=o6YHcDLod8A\nhttps ://www.youtube.com/watch?v=-uoUwGpzIL0\nhttps://www.youtube.com/watch?v=kp hYc_lvKIk&list=PLkPaq00oPRfzz9O4q06rOL2dHCEX7PQwU&index=18\nhttps://www.yo utube.com/watch?v=gvJhtBdmUi8&t=897s\n\nProfile links: \nhttps://scholar.g oogle.com/citations?user=fJp6O0UAAAAJ&hl=en\nhttps://www.linkedin.com/in/k arthika-vijayan/\nhttps://www.researchgate.net/profile/Karthika-Vijayan\n\ n### Pre - Talk Reading List\n[Few blogs/concepts to go through before the talk](https://docs.google.com/document/d/1fzf-n8lkpVk9IlmSpNKST6neaOKd02r -Sm7fXYw6nbQ/edit?tab=t.0#heading=h.5kq2oor9cxdy)\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T073256Z LOCATION:Agentic AI track (Library) - Bangalore International Centre\nBang alore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/enhancing-retrieval-in -rag-the-fused-features-way-4MGFKwLnTjXgu8RnzEtisG BEGIN:VALARM ACTION:display DESCRIPTION:Enhancing retrieval in RAG: the fused features way in Agentic AI track (Library) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Real-world lessons from multi-agent architectures DTSTART:20250719T105500Z DTEND:20250719T113000Z DTSTAMP:20260421T080241Z UID:session/SUMwbVxu7iorp4PVuvpKZi@hasgeek.com SEQUENCE:18 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Applied AI Engineering & Agentic AI track CREATED:20250702T163936Z DESCRIPTION:Multi-agent systems are gaining traction across various domain s due to their ability to adapt and operate in dynamic environments while accomplishing complex tasks. This is largely attributed to their capabilit ies in tool invocation\, context management\, planning\, and inter-agent c ollaboration. While their potential is promising\, developing sophisticate d multi-agent systems for real-world applications presents a unique set of challenges\, including accuracy\, latency\, and operational costs.\n\nIn this talk\, we will explore these challenges in depth and discuss practica l strategies to address them. Through a detailed case study of a conversat ional AI assistant designed to interact with enterprise data\, we will dem onstrate the effectiveness of multi-agent systems. We will elaborate on ho w multiple agents were orchestrated to handle diverse tasks such as perfor ming dynamic computations and retrieving information from both structured and unstructured data to derive insights. We will walk you through the les sons we learned while building a production-ready multi-agent application. \n\n**Outline of the talk**\n\n- Intro (2 mins)\n- Effectiveness of the Mu lti-agent system for dynamic business environments (3 mins)\n- Challenges faced in making it Production-ready and solutions to address them (15 mins )\n- Conclusion (5 mins)\n- QnA (5 mins)\n\n\n**Key takeaways**\n\n- Bluep rint for a production-ready multi-agent system for an enterprise applicati on \n- Insights into managing structured + unstructured data\n- Approach t o improve latency and accuracy while minimising cost\n\n**Target audience* *\n\nFolks with basic understanding of Python\, APIs\, and LLM/RAG concept s. If you’ve ever written a chatbot or SQL query—you’ll feel right a t home!\n\n**Pre-Talk Reading List**\n[Few blogs/concepts to go through be fore the talk](https://docs.google.com/document/d/1G1QrfN7-9FZ01NhNW6YHxA8 UqSuVzqyMhCn2SQZ4jRY/edit?tab=t.0#heading=h.5kq2oor9cxdy)\n\n**Speaker bio **\n\nSibin Bhaskaran is a Solution Consultant and technology lover at Sah aj AI Software. With over a decade of practical knowledge\, he strives to gain a comprehensive understanding of the entire business lifecycle\, putt ing into perspective the pivotal role that technology plays in its overall success. He has extensive knowledge across diverse domains of technology\ , spanning backend development\, mobility solutions\, robotic process auto mation\, and machine learning. GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T073353Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/navigating-real-world- challenges-in-a-production-grade-multi-agent-system-SUMwbVxu7iorp4PVuvpKZi BEGIN:VALARM ACTION:display DESCRIPTION:Real-world lessons from multi-agent architectures in Agentic A I track (Auditorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Evening break DTSTART:20250719T105500Z DTEND:20250719T113000Z DTSTAMP:20260421T080241Z UID:session/RrdmNsVxjNizXJBk9BGz4u@hasgeek.com SEQUENCE:3 CREATED:20250704T160906Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20250710T051655Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Evening break in Data Engineering and Infrastructure track (Se minar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:All’s well that ends in consistency: Change Data Capture (CDC) w ithout chaos DTSTART:20250719T113000Z DTEND:20250719T120500Z DTSTAMP:20260421T080241Z UID:session/7QVBVt6M3Q87Xkf97rFuv6@hasgeek.com SEQUENCE:13 CATEGORIES:Speaking at the Fifth Elephant 2025 Annual Conference,30 mins t alk,Data & ML Infrastructure track CREATED:20250704T155058Z DESCRIPTION:## Abstract\nChange Data Capture (CDC) is essential for real-t ime data architectures\, but when handling critical financial transactions \, even the slightest data loss or inconsistency is unacceptable. In this talk\, we’ll share our experience designing and building a highly reliab le\, low-latency CDC pipeline for a finance use case where data integrity\ , availability\, and observability were top priorities.\n\nWe’ll dive in to the key challenges we faced—ensuring exactly-once processing\, handli ng schema evolution\, mitigating network failures\, and optimizing for per formance without compromising consistency. Beyond data replication\, we’ ll cover how we implemented real-time monitoring\, alerting to proactively identify and resolve CDC failures. We’ll also discuss data reconciliati on strategies\, like audit logs\, and validation mechanisms to ensure data correctness across systems.\n\nWhether you're working with financial data or other mission-critical workloads\, this talk will provide practical in sights for building resilient\, fault-tolerant CDC pipelines with strong g uarantees for data integrity and observability.\n\n## Takeaways\n1. Deep d ive into real-world engineering challenges of implementing realtime data r eplication pipeline\n2. Learn strategies to enhance observability in CDC p ipelines using real-time monitoring\, alerting\, and metrics to proactivel y detect and resolve issues\n3. Designing end-to-end validation and reconc iliation workflows\n4. Lessons learned from operating mission-critical CDC in production\n\n## Audience\nThe talk is for\n- Platform/Data Engineers building real time data pipelines\n- DevOps Engineers managing the deploym ent and scaling of Data pipelines\n- Cloud Infra and SRE Teams responsible for performance\, availability\, and cost control\n\n\n## Bio\nSujit is a technology leader with over 14 years of experience in building large-scal e\, high-performance distributed systems. A full-stack polyglot developer\ , he specializes in Functional Programming\, Microservices\, Data Engineer ing\, and DevOps. Sujit has led complex data engineering projects\, optimi zation problems\, and cloud-native architectures\nLinkedIn: https://www.li nkedin.com/in/sujitkamthe/\n\n## Pre - Talk Reading List \n[Few blogs/reso urces to go through before the talk](https://docs.google.com/document/d/18 jpB5MeggpmlAE8kGkWcZw4AY9s6gxVlrKwk_nl9_vA/edit?tab=t.0#heading=h.5kq2oor9 cxdy)\n# Reference Links\nSlides: https://docs.google.com/presentation/d/1 zlOGIgOcxbQmVq0b85GHnIyP0XG2J__DkitfgikkPvU/edit?usp=sharing\n\nUpdated sl ides:\nhttps://docs.google.com/presentation/d/1O36m1n51xhSiZ8lDermUOaN3dga ZA57CgjuAyeqvTz0/edit?usp=sharing GEO:12.9666826;77.6352903 LAST-MODIFIED:20250712T073459Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2025/schedule/engineering-reliable-a nd-scalable-realtime-change-data-capture-pipelines-7QVBVt6M3Q87Xkf97rFuv6 BEGIN:VALARM ACTION:display DESCRIPTION:All’s well that ends in consistency: Change Data Capture (CD C) without chaos in Data Engineering and Infrastructure track (Seminar Hal l) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Wrap up & Feedback | Agentic AI Track DTSTART:20250719T113000Z DTEND:20250719T115000Z DTSTAMP:20260421T080241Z UID:session/2aXYxKAfKCwcuSgiNKDBtU@hasgeek.com SEQUENCE:8 CREATED:20250710T122040Z DESCRIPTION:Venue: Ground floor Foyer (outside Auditorium) GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T074418Z LOCATION:Agentic AI track (Auditorium) - Bangalore International Centre\nB angalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Wrap up & Feedback | Agentic AI Track in Agentic AI track (Aud itorium) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Wrap up & Feedback | Data Engineering and Infrastructure Track DTSTART:20250719T120500Z DTEND:20250719T122500Z DTSTAMP:20260421T080241Z UID:session/SrqikiWWvGhkAPmovgz7km@hasgeek.com SEQUENCE:9 CREATED:20250704T160601Z DESCRIPTION:Venue: First floor Foyer (outside Seminar Hall) GEO:12.9666826;77.6352903 LAST-MODIFIED:20250717T074427Z LOCATION:Data Engineering and Infrastructure track (Seminar Hall) - Bangal ore International Centre\nBangalore\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Wrap up & Feedback | Data Engineering and Infrastructure Track in Data Engineering and Infrastructure track (Seminar Hall) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT END:VCALENDAR