BEGIN:VCALENDAR VERSION:2.0 PRODID:-//HasGeek//NONSGML Funnel//EN DESCRIPTION:On AI\, industrial applications of ML\, and MLOps X-WR-CALDESC:On AI\, industrial applications of ML\, and MLOps NAME:The Fifth Elephant 2023 Monsoon X-WR-CALNAME:The Fifth Elephant 2023 Monsoon REFRESH-INTERVAL;VALUE=DURATION:PT12H SUMMARY:The Fifth Elephant 2023 Monsoon TIMEZONE-ID:Asia/Kolkata X-PUBLISHED-TTL:PT12H X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT SUMMARY:Check-in DTSTART:20230811T033000Z DTEND:20230811T042000Z DTSTAMP:20260421T120126Z UID:session/GfJncx1eUMx6Xa9DmXp8nJ@hasgeek.com SEQUENCE:2 CREATED:20230709T165311Z LAST-MODIFIED:20230709T165420Z LOCATION:Bangalore ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Check-in in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Introduction to The Fifth Elephant 2023 DTSTART:20230811T042000Z DTEND:20230811T043000Z DTSTAMP:20260421T120126Z UID:session/X8PGFo1vfFWgw35p1R2P7f@hasgeek.com SEQUENCE:1 CREATED:20230709T165448Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230709T165452Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Introduction to The Fifth Elephant 2023 in Featured talks (Aud itorium - first floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Time Series: A Tale of Three Generations of Forecasting Models DTSTART:20230811T043000Z DTEND:20230811T051500Z DTSTAMP:20260421T120126Z UID:session/LCJRwoavYFV4Hb86JdQ9un@hasgeek.com SEQUENCE:4 CATEGORIES:Confirmed,AI and research CREATED:20230709T173040Z DESCRIPTION:## Link to presentation: \n\nhttps://drive.google.com/file/d/1 hH90FCWxRFv0IQcoBqX2yRM1DcTzDBQr/view?usp=sharing\n\n\n## Abstract: \n\nTh is talk addresses time-series modeling for demand forecasting\, covering a brief history of various types of forecasting models and showcasing where each class of methods is applicable. The talk also covers a case study wi th real-world data from Sortly (an inventory management SAAS company)\, hi ghlighting the challenges faced with diverse consumption patterns and anom alies. Specifically\, the results of forecasting with SARIMAX\, Random For ests and a Transformer based model are discussed for this case-study. Addi tionally\, the talk shares some behind the scenes prompt engineering stori es for this case-study - including some detailed prompts\, what worked and what did not.\n\n## What will you get out of this talk:\n\nBy attending t his talk\, attendees will gain insights into various types of time-series models and their effectiveness in demand forecasting for various types of data. \n \n\n## Talk Outline: \n\nPart 1: A brief History of Demand Foreca sting: \nVarious models for demand forecasting\nTypes of Models\nDecomposa ble Time Series Models - Ex: Arima\, Sarimax\, Prophet\nClassical ML and E nsemble based approaches - Ex: Random Forest\, LightGBM\nDeep learning bas ed models - Example: LSTMs\, GRU based model\, Transformers\nWhere do each of these models make sense to use\n\nPart 2: Demand Forecasting for Sortl y \nUnderstanding Sortly Data\nChallenges - Anomalies\, diverse workflows\ nComparison of models for select data\n\nPart 3: Prompt Engineering Storie s\nWhere all can one use prompt engineering\nExamples of detailed prompts\ nExamples of problems encountered with prompts\nPrompting Lessons GEO:12.9666826;77.6352903 LAST-MODIFIED:20230915T125114Z LOCATION:AI and Research track (Library - 2nd floor) - Bangalore Internati onal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/time-series-modelli ng-for-demand-forecasting-LCJRwoavYFV4Hb86JdQ9un BEGIN:VALARM ACTION:display DESCRIPTION:Time Series: A Tale of Three Generations of Forecasting Models in AI and Research track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:The data bottleneck in distributed AI/ML workloads DTSTART:20230811T043000Z DTEND:20230811T051500Z DTSTAMP:20260421T120126Z UID:session/AVQGVXY5YK9gEc2ZrYeVgo@hasgeek.com SEQUENCE:20 CREATED:20230709T173643Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T023938Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:The data bottleneck in distributed AI/ML workloads in Industri al track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Jugalbandi.ai: Powering breakthrough Conversation AI for every Ind ian DTSTART:20230811T043000Z DTEND:20230811T051500Z DTSTAMP:20260421T120126Z UID:session/JQUbC3xgNpvqNYJUihPdVS@hasgeek.com SEQUENCE:7 CATEGORIES:Confirmed CREATED:20230709T165410Z DESCRIPTION:### Introduction\nThe gap in access to Justice in India is a h uge opportunity for AI. Initial field trials from Jugalbandi with farmers\ , domestic workers and waste pickers presents several lessons in AI produc t design for all Indians. In this talk we will gain a deeper understanding of the opportunities in addressing access to justice\, the design choices in building Jugalbandi and what kind of technical challenges come in buil ding these products.\n\n\n### Part 1 (15 min)\nLandscape of AI for Justice Opportunities. We start by getting a perspective on the field of Justice and how the text of Law is a subset of the field of Justice administration s.\n\n\n![Migrants who are saying to the judge over ODR session that they are okay with 2 lakh rupees](https://images.hasgeek.com/embed/file/ab27af6 2dd8c49c99c99e082413024b0)\n\n*This image is not victory as perceived by m ost people\, but migrants who are saying to the judge over Online Dispute Resolution (ODR) session that we are okay with 2 lakh rupees.*\n\nThe curr ent state of access to justice is very low where only 1 in 10 people with a legal need reaches the formal system. One common theme across access to justice is being able to find legal information or legal services at a low cost in vernacular. We end this section by getting a sense of where conve rsational AI could be useful for the access to justice.\n\n### Part 2 (20 min)\nWe talk about the building blocks of Jugalbandi - a reasoning engine \, voice and translation models and keeping the AI truthful. We show how t he system is connected to WhatsApp and what kind of privacy related challe nges we face and what are some of the measures we are taking.\n\n### Part 3 (15 min)\nWe end the talk by sharing challenges in scaling these systems \, the opportunities to create autonomous agents to help citizens. GEO:12.9666826;77.6352903 LAST-MODIFIED:20240425T055954Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/jugalbandi-ai-power ing-breakthrough-conversation-ai-for-every-indian-JQUbC3xgNpvqNYJUihPdVS BEGIN:VALARM ACTION:display DESCRIPTION:Jugalbandi.ai: Powering breakthrough Conversation AI for every Indian in Featured talks (Auditorium - first floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Elevating Model Training at DoorDash with Ray DTSTART:20230811T043000Z DTEND:20230811T051500Z DTSTAMP:20260421T120126Z UID:session/49nvPmx3nVr5jpzfaERcVn@hasgeek.com SEQUENCE:4 CATEGORIES:Confirmed,AI and Product CREATED:20230709T171506Z DESCRIPTION:At DoorDash\, machine learning is a key component\, used to en hance the experience of merchants\, dashers\, and customers. As our machin e learning use cases keep growing\, our forecasting and training pipelines are faced with several challenges like scalability\, growing costs\, redu ced user development velocity and lack of proper debugging/observability.\ n\nDriven by these challenges\, we started learning about Ray and implemen ted a POC to verify the feasibility of incorporating Ray into our existing ML Platform architecture. In this talk\, we’ll share our journey toward s adopting Ray\, working with the open source KubeRay\, how we built our P OC and benchmarking setup and also share what the future of ML Platform at DD looks like with Ray being a core component. GEO:12.9666826;77.6352903 LAST-MODIFIED:20230907T111914Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/elevating-model-tra ining-at-doordash-with-ray-49nvPmx3nVr5jpzfaERcVn BEGIN:VALARM ACTION:display DESCRIPTION:Elevating Model Training at DoorDash with Ray in AI and Produc t track\, Lightning talks and BOFs (Seminar halls - 1st floor) in 5 minute s TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Adaptive Metric Alignment for Demand Forecasting in Swiggy Instama rt DTSTART:20230811T051500Z DTEND:20230811T060000Z DTSTAMP:20260421T120126Z UID:session/6pn5eYVYL61H81b1NXX88d@hasgeek.com SEQUENCE:8 CATEGORIES:Confirmed,AI and research CREATED:20230709T172055Z DESCRIPTION:Problem statement:\nInstamart\, the quick commerce grocery del ivery service of Swiggy gives unparalleled convenience of being able to or der\, from a huge assortment\, across fresh fruits & vegetables/ dairy/FMC G products and accessories for household requirements\, parties or festivi ties\, pretty much at any time of the day and also through late night (fro m 6am to 3pm) and get the delivery in ~10–15 min. Instamart follows a da rk store model where micro-fulfilment centers are established to fulfill t he grocery orders of a certain geographical area of a few kilometers of ra dius. Efficient demand planning ensures that the sufficient units of each of the products are ‘available’ in the closest pod (dark store) for cu stomers to order throughout the day\, while making sure not stocking up to o many units which can eventually lead to ‘wastage’. For efficient pla nning\, ML based forecasting techniques are used to predict the daily ‘d emand’ of an item for a given store (referred as SKU). But the demand fo recasting for Instamart\, or instant grocery delivery systems in general\, have a handful of challenges that traditional forecasting methodologies c an not resolve.\n\nFirstly\, due to the hyper-local nature of demand plann ing\, there is high variation of demand across geographies\, items and day s – which leads to frequent ‘out-of-stock’ for some of the SKUs even before the pod is closed for the day. Hence\, the historically observed t ime series data for building forecasting models is not the accurate repres entation of the ‘true’ demand\, rather it is a truncated demand. The f requent absence of true demand makes the model development and evaluation challenging\, especially when we are dealing with number of SKUs in the or der of 10^4.\n\nBusiness implication:\nTo track the efficiency of the dema nd planning\, the business team tracks two metrics primarily: 1) availabil ity – it measures the proportion of the day a SKU was available for the users to order\, and 2) wastage – it approximately quantifies the units over-stocked and eventually led to wastage. Not being able to accurately e valuate the model performance using traditional metrics such as wMAPE on b ack testing data can lead to deploying forecasting models in production th at can either underpredict or overpredict the true demand which means lowe r availability (i.e.\, revenue opportunity loss) or higher wastage respect ively.\n\nSolution:\nIn this presentation we will go over our approach of ‘Adaptive Metric Alignment’ for accurate model evaluation which is mor e closely aligned with the business metrics. We will cover the following t opics in our presentation:\n1. Introduction to demand planning process for Swiggy Instamart\n2. Demand forecasting challenges for quick commerce gro cery delivery services\n3. Drawbacks of traditional model evaluation metho ds\n4. Adaptive metric alignment: estimated availability and wastage\nProd uction implication and conclusion\n\n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20240425T060006Z LOCATION:AI and Research track (Library - 2nd floor) - Bangalore Internati onal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/adaptive-metric-ali gnment-for-demand-forecasting-in-swiggy-instamart-6pn5eYVYL61H81b1NXX88d BEGIN:VALARM ACTION:display DESCRIPTION:Adaptive Metric Alignment for Demand Forecasting in Swiggy Ins tamart in AI and Research track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:AI Founder Panel: LLMs\, Moats\, Margins and Funding DTSTART:20230811T051500Z DTEND:20230811T060000Z DTSTAMP:20260421T120126Z UID:session/GettF1Lrtf8AS9mE3mXixe@hasgeek.com SEQUENCE:11 CREATED:20230709T172224Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230908T114724Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:AI Founder Panel: LLMs\, Moats\, Margins and Funding in AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Bumpy Roads\, High Speeds: My Unexpected Journey from PhD to Tech Entrepreneurship DTSTART:20230811T051500Z DTEND:20230811T060000Z DTSTAMP:20260421T120126Z UID:session/HPyyTJ8kRzNhQ8vRKVFV2a@hasgeek.com SEQUENCE:9 CATEGORIES:Confirmed CREATED:20230709T165510Z DESCRIPTION:### Background\nIn 2015\, I sold the intellectual property (IP ) of my Silicon Valley company\, Perceptive Code LLC\, to Mercedes Benz. S ubsequently\, I was tasked with meeting certain milestones as part of the handover process. I chose to complete this task in India\, where I aways w anted to be. I successfully downsized our research model\, initially consu ming 6GB of GPU memory\, to a mere 300KB of weights. I then implemented th is model on a 2W FPGA with 112 DSP blocks. This model is now included as p art of the MBUX option in all Mercedes cars. Using cameras installed in th e car\, it enables gesture control for various functions such as turning o n passenger side lamps when a hand is extended to retrieve something\, or choosing which rearview mirror to adjust simply by looking at it\, among o ther features.\n\n### What will this talk cover?\nThis talk will be a pers onal reflection on the past decade of my life. I'll discuss my journey fro m leaving my job at Yahoo! in Bangalore and moving to Italy\, then Germany for my PhD. I will share how I didn't secure the jobs I initially aimed f or post-PhD and subsequently "settled" for a post-doctoral position with C hris Bregler at the Courant Institute at NYU. Intriguingly\, I unintention ally found myself working with Yann LeCun\, which led to the publication o f four papers alongside this Turing Award winner.\n\nI'll touch upon the p re-AlexNet era and provide insights into the academic environment at NYU d uring that time. I'll also delve into my experience with Apple's self-driv ing team and discuss my decision to leave Apple for a riskier venture with Mercedes\, striving to meet milestones and deliver a product. I'll discus s how\, in hindsight\, this all unfolded.\n\nAdditionally\, I plan to shar e my experiences in academia. Despite having offers and spending time at p restigious institutions like IIT Bombay and IISc\, I made the choice not t o pursue academia full-time. Instead\, I continued to explore new ventures such as UAVIO Labs and Fastcode AI.\n\n### Talk Outline\n- My PhD journey : Machine Learning for Computer Graphics\, spending time at Weta\, earning credits for Adventures of Tintin\, 2011\n- The jobs I desired but didn't secure after my PhD\n- My tenure at NYU: people at NYU at that time\, wor king with Yann\, working on Theano\, and projects I undertook\n- How and w hen Jonathan and I started Perceptive Code LLC\n- Life after my post-doc: why I chose Apple\n- My conversations with Mercedes and the ensuing deal\, coming back to India\n- The efforts behind building the product: technolo gy\, papers\, and patents produced\n- My current aspirations and what I lo ok forward to GEO:12.9666826;77.6352903 LAST-MODIFIED:20240425T060024Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/bumpy-roads-high-sp eeds-my-unexpected-journey-from-phd-to-tech-entrepreneurship-HPyyTJ8kRzNhQ 8vRKVFV2a BEGIN:VALARM ACTION:display DESCRIPTION:Bumpy Roads\, High Speeds: My Unexpected Journey from PhD to T ech Entrepreneurship in Featured talks (Auditorium - first floor) in 5 min utes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Learnings from Building Deep Learning Models for Better Cardiac Ca re DTSTART:20230811T051500Z DTEND:20230811T060000Z DTSTAMP:20260421T120126Z UID:session/N4MxgU8bwFBJkjDzx91SZg@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed CREATED:20230709T172508Z DESCRIPTION:## Abstract: \nEchocardiogram(Echo) is one of the common modal ity that captures state of the heart in the form images and videos. Using Ultrasound technique\, an echo study captures multiple cross sections of t he heart\, termed as Views. Cardiologist utilizes measurements on the basi s of these Views to analyse heart functions. In order to automate this pro cess to measure how heart is functioning\, an initial step is to identify Echo View types\, akin to image classification. \n\nIn this talk\, we will present development of Auto View Classification. We will emphasize on dat aset development for Echo Views\, challenges of Inter-Observer-Agreement\, choice of deep learning models and showcase early results comparable with state-of-the-art. Finally\, we will provide general guidelines while buil ding AI models for healthcare in general. \n\n## Outline:\n- Layman's in tro to Heart and Echo\n- What is a View and why it is important for Echo d iagnosis.\n- Why automatically identification of Echo View is needed?\n- O verview of Auto View Classification with brief on Dataset development\, i nter-obeserver agreement\, choice of Deep learning and SOTA results\n- Lea rnings from building vision models in cardiology.\n \n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20240425T060029Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/learnings-from-buil ding-deep-learning-models-for-better-cardiac-care-N4MxgU8bwFBJkjDzx91SZg BEGIN:VALARM ACTION:display DESCRIPTION:Learnings from Building Deep Learning Models for Better Cardia c Care in Industrial track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Break DTSTART:20230811T060000Z DTEND:20230811T063000Z DTSTAMP:20260421T120126Z UID:session/FF2JC6yb6CLxw7s7CsbxZu@hasgeek.com SEQUENCE:1 CREATED:20230709T165526Z LAST-MODIFIED:20230709T165813Z LOCATION:Bangalore ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Break in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:A Customised Speech Recognition System for the Indian Telesales In dustry DTSTART:20230811T063000Z DTEND:20230811T071500Z DTSTAMP:20260421T120126Z UID:session/NCRodD5nT9fUpUWN2LJcGW@hasgeek.com SEQUENCE:5 CREATED:20230730T174012Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T024417Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:A Customised Speech Recognition System for the Indian Telesale s Industry in Industrial track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Solving for explainability of fraud detection models DTSTART:20230811T063000Z DTEND:20230811T071500Z DTSTAMP:20260421T120126Z UID:session/X8QGNUFJaexSwppSWDbJ4R@hasgeek.com SEQUENCE:5 CATEGORIES:Confirmed CREATED:20230709T173404Z DESCRIPTION:Problem : \nAt the TnS(Trust and Safety) team at Swiggy\, buil ding powerful fraud detection models that operate at high precision while still capturing maximum fraud has been the uber goal. Our system currently operates at a high level of complexity through various interventions\, mo delling techniques\, and semi-supervised training methods while maintainin g robustness.\nFor the final downstream model\, we have always relied on t ree-based learners over neural networks. Since are data is primarily tabul ar in nature\, tree-based learners outperformed DNNs significantly on the winning metrics. While tree-based learners are great performers in terms o f the final metrics that we're looking to optimise\, it has a few challeng es:\n 1. It inherently restricts us from trying out more complex da ta structures like images or sequential data\, we have tried to integrate such signals through a separate model whose final score is fed into the tr ee based learner but it significantly adds to complexity of the system.\n 2. A major press point for Fraud models historically has been a lac k of explainability in predictions. We have experimented with LIME and SHA P-based approaches to build an explainable overhead but they’re computat ionally expensive to run for each record.\n\nSolution:\nWhile tree-based m ethods for a deployable model have all these challenges\, what works in th eir favour is that they have historically outperformed DL-based methods by a significant margin. This changes with TabNet\, in the original paper(Re f)\, authors claim that TabNet can match or even outperform tree-based met hods while also giving sample-level explainability\, which we can also vis ualise. We explored a tabnet based model for our approach and found it to be on par with tree-based counterpart(xgboost). TabNet also allowed us to compute and store feature level attention within the model logs without an y computational overhead.\n\nOutline:\nIn the presentation\, we'll be goin g through the following in depth.\nCurrent pipeline and solution\nChalleng es in depth\nMotivation for TabNet and what it unlocks\nExperimental resul ts and conclusion GEO:12.9666826;77.6352903 LAST-MODIFIED:20230907T112328Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/solving-for-explain ability-of-fraud-detection-models-X8QGNUFJaexSwppSWDbJ4R BEGIN:VALARM ACTION:display DESCRIPTION:Solving for explainability of fraud detection models in AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Vigil: Effective end-to-end monitoring for large-scale recommender systems at Glance DTSTART:20230811T063000Z DTEND:20230811T071500Z DTSTAMP:20260421T120126Z UID:session/3o2y5E1gtJmtVyM7proBCp@hasgeek.com SEQUENCE:4 CATEGORIES:Shortlisted,Innovative work CREATED:20230709T165803Z DESCRIPTION:\n## Abstract\nThe success of large-scale recommender systems hinges upon their ability to deliver accurate and timely recommendations t o a diverse user base. At Glance\, we offer snackable personalized content to the lock screens of 200M smartphones. In this context\, continuous mon itoring is paramount as it safeguards data integrity\, detects drifts\, ad dresses evolving user preferences\, optimizes system downtime\, and ultima tely augments the system's effectiveness and user satisfaction. This talk explores the critical role of continuous monitoring in our ecosystem. We i ntroduce Vigil\, a comprehensive end-to-end monitoring framework designed specifically for Glance's recommender systems. These practices revolve aro und three key pillars: mitigating developer fatigue\, ensuring precise pre dictions\, and establishing a centralized monitoring framework. By adoptin g these practices\, we have observed an 18% increase in user engagement\, a 30% reduction in compute cost\, a 26% drop in downtime\, and a surge in developer productivity demonstrated by a 45% decrease in turnaround time.\ n\n\n## Impact of Vigil\n- Implementing a centralized system monitoring vi ew has led to enhancements in developer productivity\, effectively mitigat ing alert fatigue and reducing turnaround time to detect and fix issues by 45% during on-call operations.\n- Proactively identifying drifts in data and models has yielded more robust and precise predictions\, resulting in an 18% increase in engagement metrics.\n- By effectively monitoring latenc ies\, errors\, and resource utilization\, coupled with practices such as a daptive retraining and eliminating redundant data and pipelines\, we have achieved a commendable 30% reduction in system costs\, further optimizing performance and resource management.\n- Additionally\, there has been a 26 % decrease in system downtime\, thus ensuring enhanced reliability and uni nterrupted service for our users.\n\n\n## Key takeaways from the talk\nImp lementing Vigil has led to tangible improvements in key performance metric s\, showcasing the value of effective end-to-end monitoring in large-scale recommender systems. Glance's experience with Vigil highlights the import ance of continuous monitoring. The talk offers valuable insights that can be applied to similar large-scale recommender systems\, benefiting system performance\, user engagement\, cost-efficiency\, and developer productivi ty.\n\n## Outline of the talk\n- Introduction to Glance\n- Challenges in m onitoring large-scale recommender systems\n- Vigil: A comprehensive end-to -end monitoring framework/practices\n - Proactive alerting and system mo nitoring\n - Dependency and impact monitoring\n - Testing and performa nce monitoring\n- Impact of adopting Vigil at Glance\n- Expressway: A cent ralized monitoring tool built on the ideas of Vigil\n- Key Takeaways\n\n** Link to Full Talk Proposal** - [Vigil: Effective end-to-end monitoring for large-scale recommender systems at Glance](https://drive.google.com/file/ d/1rXyiytOBPyRPf2wCiTl0sOeIcqjhNvLD/view?usp=sharing)\n\n\n### Labels\nMLO ps\, Recommender Systems\, ML Model lifecycle\, ML Monitoring Best Practic es\, ML Monitoring Implementation\n\n## Speakers\n- [Priyansh Saxena](http s://www.linkedin.com/in/saxenapriyansh/)\, Data Scientist\, InMobi Group\n - [Manisha R](https://www.linkedin.com/in/r-manisha/)\, Data Scientist\, I nMobi Group\n\n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20230825T131345Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/vigil-effective-end -to-end-monitoring-for-large-scale-recommender-systems-at-glance-3o2y5E1gt JmtVyM7proBCp BEGIN:VALARM ACTION:display DESCRIPTION:Vigil: Effective end-to-end monitoring for large-scale recomme nder systems at Glance in Featured talks (Auditorium - first floor) in 5 m inutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Building real-time video-based analytics product DTSTART:20230811T063000Z DTEND:20230811T071500Z DTSTAMP:20260421T120126Z UID:session/Lm9e6FMTKvqdBseFTMGHxA@hasgeek.com SEQUENCE:4 CATEGORIES:Confirmed,AI and Product CREATED:20230709T171923Z DESCRIPTION:# Abstract:\n\nCompanies are now building products using real- time computer vision and machine learning of video from various systems an d processes. This requires a full stack system consisting of video ingesti on and storage\, live inferencing\, post-processing of data generated\, an d use of the data in the business context or customer’s domain language. We demonstrate a preferred architecture and stack to build such a system\ , the challenges & the design choices and shall have elements useful to an yone building video analytics-based systems.\n\n# Description\nWe’ll tal k about some possible architectures for building a real-time video analyti cs inferencing system. We’ll cover:\n- Problem\, use cases - ie object o r action detection & recognition. What makes this problem different & more challenging.\n- Various high level industry approaches to them - right fr om 3D Conv\, 2 stream spatio-temporal approaches\, proposal generation to recent transformer based approaches.\n- Video infra - preferred protocol c hoices for video ingestion\, storage formats\, libraries to process stream ing video\, streaming architectures to infer on them\, \n- Translating the model output into the customer context product\, modeling customer domain & common frameworks for them\, and converting the data into insights. \n- We’ll also touch on some of the scalability and implementation challeng es of this stack\, identifying bottlenecks\, prioritizing them and designi ng options to solve for scalability.\n\nThe talk leverages some of the spe aker’s prior experience building scalable systems at Google\, building B 2B applications\, and Drishti.\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20230915T130247Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/title-building-real -time-video-based-analytics-product-Lm9e6FMTKvqdBseFTMGHxA BEGIN:VALARM ACTION:display DESCRIPTION:Building real-time video-based analytics product in AI and Pro duct track\, Lightning talks and BOFs (Seminar halls - 1st floor) in 5 min utes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transforming COD from a Risk to Growth lever using Machine Learnin g DTSTART:20230811T071500Z DTEND:20230811T080000Z DTSTAMP:20260421T120126Z UID:session/PZzAwsqUKMjiLczk9z1rv2@hasgeek.com SEQUENCE:14 CREATED:20230709T172543Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T023514Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transforming COD from a Risk to Growth lever using Machine Lea rning in Industrial track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Leadership Panel: Building Teams\, Research & Products DTSTART:20230811T071500Z DTEND:20230811T080000Z DTSTAMP:20260421T120126Z UID:session/8funFTJvt1MoCzwpQiL1b8@hasgeek.com SEQUENCE:6 CREATED:20230709T165839Z DESCRIPTION:Panelists: Goda Ramkumar (VP\, Data Science\, Swiggy)\, Harsh Singhal (Head of ML & AI\, Koo)\, Sumod Mohan (Founder\, CEO\, Autoinfer) & Arvind Saraf (Founder\, Attention Tag). \n\nHow did the current leaders in the ML/AI space find themselves here? What were those moments that conv inced them to change their career trajectory into data science? How do the y pick up newer areas to venture into & how they collaborate across teams in their organisations? GEO:12.9666826;77.6352903 LAST-MODIFIED:20230907T113055Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Leadership Panel: Building Teams\, Research & Products in Feat ured talks (Auditorium - first floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:7 mistakes while deploying GenAI apps in production DTSTART:20230811T071500Z DTEND:20230811T074000Z DTSTAMP:20260421T120126Z UID:session/Qnd5jx1BynoL4AWvkYnEQR@hasgeek.com SEQUENCE:12 CATEGORIES:Confirmed,Lightning talks CREATED:20230709T173854Z DESCRIPTION:I've built production LLM systems and made multiple mistakes. The talk will focus on key areas to focus on while deploying LLM and Gener ative AI systems in production GEO:12.9666826;77.6352903 LAST-MODIFIED:20230908T114822Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/7-mistakes-while-de ploying-genai-apps-in-production-Qnd5jx1BynoL4AWvkYnEQR BEGIN:VALARM ACTION:display DESCRIPTION:7 mistakes while deploying GenAI apps in production in AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Demystifying Quantisation in Large Language Models in Plain Englis h with Basic Math DTSTART:20230811T071500Z DTEND:20230811T080000Z DTSTAMP:20260421T120126Z UID:session/36xFURMK1upkY5FSzh3Vz3@hasgeek.com SEQUENCE:7 CATEGORIES:Confirmed,AI and research CREATED:20230709T173245Z DESCRIPTION:# Demystifying Quantisation in Large Language Models in Plain English with Basic Math\n## What will I cover\n- I will cover some basic m aths of sizing up the memory and compute requirements of training and infe rence of a large language model. Some popular open source models will be u sed as example.\n- Quick brush up of data types.\n- In plain English\, how are popular quantisation methods working?\n- Take an example of a typical computation in a neural network and show what quantisation brings to the table.\n- What impact does this make on compute\, memory requirements? Wha t is the fine print?\n- Why is this important? How can you apply this in y our work?\n\n## Why is this topic is important?\nQuantisation has emerged as a significant enabler for large language models (LLMs)\, making them ac cessible for companies without extravagant budgets (read: throw money at t he problem) and paving the way for edge deployments. This talk delves beyo nd the basic concept of converting floats to integers. I'll explain the un derlying math that governs the memory and computation requirements\, demon strating how quantisation computations facilitate not only inference but a lso\, potentially\, training. Additionally\, I will illuminate the cost\, computational\, and business impacts of quantisation.\n\n## What can audie nce learn from it?\n- Intuitive yet in-depth comprehension of why quantisa tion is crucial for training or fine-tuning LLMs.\n- What is\, roughly\, h appening in the maths? Where are the trade-offs?\n- How does it impact acc uracy? What is the evidence for its claims?\n- How to make informed quanti sation trade-offs\, equipping them to exploit LLMs across various use case s effectively.\n\n## Past Experience\nI have in past hosted many talks on ML/AI at Fifth Elephant and other conferences. These includes hands on wor kshops and short talks. I’m obsessed about giving a clear understanding of underlying maths fundamentals while also explain the business impact. \ n\n## Profile Links\n- [LinkedIn](https://www.linkedin.com/in/harshadss/?o riginalSubdomain=in)\n- [Gave a talk recently at a Bengaluru meetup on som ewhat similar topic](https://www.meetup.com/bangalore-genai-meetup/events/ 294133924)\n- [Instructor at London School of Economics Journalism AI acad emy for 2023](https://blogs.lse.ac.uk/polis/2023/06/06/introducing-the-202 3-journalismai-academy-emea-and-apac-cohorts/)\n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20240207T084503Z LOCATION:AI and Research track (Library - 2nd floor) - Bangalore Internati onal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/demystifying-quanti sation-in-large-language-models-in-plain-english-with-basic-math-36xFURMK1 upkY5FSzh3Vz3 BEGIN:VALARM ACTION:display DESCRIPTION:Demystifying Quantisation in Large Language Models in Plain En glish with Basic Math in AI and Research track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Algorithm Friendly Design: Aligning product design with algorithmi c requirements DTSTART:20230811T074000Z DTEND:20230811T080000Z DTSTAMP:20260421T120126Z UID:session/7pKc3Kj2gf3iXmmRBNppUs@hasgeek.com SEQUENCE:7 CREATED:20230801T151206Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230908T115229Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Algorithm Friendly Design: Aligning product design with algori thmic requirements in AI and Product track\, Lightning talks and BOFs (Sem inar halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch DTSTART:20230811T080000Z DTEND:20230811T090000Z DTSTAMP:20260421T120126Z UID:session/LsSoUGTqW4t2RA9nGiNb8Z@hasgeek.com SEQUENCE:1 CREATED:20230709T171218Z LAST-MODIFIED:20230709T171223Z LOCATION:Bangalore ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Lunch in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Birds of Feather Session DTSTART:20230811T090000Z DTEND:20230811T094500Z DTSTAMP:20260421T120126Z UID:session/TcYbhPoHAmDFbTuNEV6tkP@hasgeek.com SEQUENCE:0 CREATED:20230809T121001Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230809T121001Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Birds of Feather Session in Featured talks (Auditorium - first floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Forecasting Architecture\, Metrics\, and Learnings @ Samsung Ads DTSTART:20230811T090000Z DTEND:20230811T094500Z DTSTAMP:20260421T120126Z UID:session/KknmrHg74qp4cckdiGAKGj@hasgeek.com SEQUENCE:14 CATEGORIES:Shortlisted,Not applicable,Confirmed,Data Science CREATED:20230709T171528Z DESCRIPTION:Samsung Ads is an intuitive audience platform that delivers me aningful experiences reaching the right audience across screens\, formats and devices. With more than 900M Mobiles and 150M Smart TVs\, and the larg est first party data set powered by ACR\, we help marketers reach targets and enhance experiences that span digital landscapes. The business has gro wn 10x since 2015. Our foundation is based on Samsung’s strength as a ma nufacturer in two key connected consumer device spaces: Mobiles & Smart TV s ...from which we derive two critical components that power our Samsung A ds businesses today: Data and Ad Impressions. We combine these assets to c reate powerful Ad offerings that drive reach\, performance\, and return on Ad spend for the world’s leading marketers.\nSamsung Ads needs to sell to advertisers in advance to show advertisements. The Ad opportunities dep end on user behaviour – users turning the TV on and going to specific sc reens. Having an automated way to forecast availability opportunities is c ritical to:\n• Ensure we do not over-commit to advertisers (monetary imp lications + hurts reputation)\n• Ensure we do not under-commit (opportun ities are wasted\, potential revenue wasted)\n• Ensure users are able to self service the forecast process (tap into a larger market segment that wants self service\nOur goal is to predict how many impression events will be received for a specific campaign over the duration of the campaign. Th e campaigns are setup to target opportunities based on various criteria in cluding Location\, Time of day\, TV Model\, Type of ad opportunity and Use r identifiers.\nForecasting is a complex problem that typically involves a single time series\, and to predict for one step ahead. We have many diff erent time series patterns\, and multi-step forecasts with a long forecast horizon (over 90 days). This requires the use of sequence to sequence mod els and modern techniques such as transformer architectures. We are buildi ng this solution using the state of the art Temporal Fusion Transformer mo dels. We will go over the different type of Ads such as Roadblocks\, Audie nce Take overs\, Rotationals and Video Ads and the factors affecting forec asting.\nWe will go over the key challenges faced in coming up with a work ing model architecture\, such as erroneous ground truth data\, data availa bility & quality issues and data understanding gaps\, and our approaches t o deal with these challenges. We will go over the use of data sketches and an OLAP DB like druid to get past data and use that and other features as inputs to a TFT model. For modeling external competition\, we will explai n how we estimate price dependence of win rates using survival models. We will also introduce the evaluation framework built to evaluate the forecas ting accuracy at three different phases - during development\, pre-releas e and post release. Earlier\, the analysis was done manually which had man y challenges like lack of consistency\, delays\, lack of historical data e tc. which were solved with the evaluation framework.\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20231123T114730Z LOCATION:AI and Research track (Library - 2nd floor) - Bangalore Internati onal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/forecasting-samsung -ads-KknmrHg74qp4cckdiGAKGj BEGIN:VALARM ACTION:display DESCRIPTION:Forecasting Architecture\, Metrics\, and Learnings @ Samsung A ds in AI and Research track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Synthetic Sorcery: Fooling Neural Networks with Unreal Data for Re al-World Applications DTSTART:20230811T090000Z DTEND:20230811T094500Z DTSTAMP:20260421T120126Z UID:session/UHVL7xGHMCzJHFKsy9oaDi@hasgeek.com SEQUENCE:4 CATEGORIES:Confirmed,Industrial CREATED:20230709T172851Z DESCRIPTION:The integration of computer vision technologies in the automot ive industry has revolutionized various aspects of vehicle safety\, naviga tion\, and driver assistance systems. However\, developing robust and accu rate computer vision models for real-world scenarios necessitates large-sc ale\, diverse\, and accurately labeled datasets\, which can be challenging to obtain. Consider a scenario where the duration of data collection and annotation efforts\, spanning from 60 to 90 days and involving hundreds of manual annotators\, is condensed to a mere 2 hours of effort resulting in impeccable\, error-free data.\n\nDrawing on our group’s experience at _ Mercedes-Benz R&D India_ in research as well as delivering customer-ready products in the area of intelligent interior and *in-cabin* sensing (MBUX interior assist)\, this talk will explore the benefits of using synthetic data to augment or replace traditional\, labor-intensive data collection m ethods.\n\nWe will delve into the techniques and methodologies employed fo r generating synthetic data tailored to computer vision tasks. This includ es modeling virtual environments\, capturing complex actions\, and incorpo rating accurate ground truth annotations. \n\nFurthermore\, we will examin e the key challenges involved in using synthetic data that closely matches real-world data distributions and the strategies used to bridge the domai n gap between synthetic and real data. This talk will highlight successful applications of synthetic data in automotive domain. Specifically\, we wi ll focus on the techniques we employed on datasets like [cityscapes](https ://www.cityscapes-dataset.com/)\, achieving *state-of-the-art* results (T op-1 when submitted\, now Top-3).\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T023822Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/synthetic-sorcery-f ooling-neural-networks-with-unreal-data-for-real-world-applications-UHVL7x GHMCzJHFKsy9oaDi BEGIN:VALARM ACTION:display DESCRIPTION:Synthetic Sorcery: Fooling Neural Networks with Unreal Data fo r Real-World Applications in Industrial track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Reinforcement Learning: From Games to chatGPT and Beyond DTSTART:20230811T090000Z DTEND:20230811T094500Z DTSTAMP:20260421T120126Z UID:session/TyxzEA9LXm1Z8LNHKua4NR@hasgeek.com SEQUENCE:8 CREATED:20230709T173636Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230915T130107Z LOCATION:AI and Research track (Library - 2nd floor) - Bangalore Internati onal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Reinforcement Learning: From Games to chatGPT and Beyond in AI and Research track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Harmonising Art and AI: Crafting Jazzy and Juicy Video Snippets th rough AI DTSTART:20230811T094500Z DTEND:20230811T103000Z DTSTAMP:20260421T120126Z UID:session/D8HzDhgfXsikWdcWLNifCC@hasgeek.com SEQUENCE:10 CREATED:20230709T171144Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T050327Z LOCATION:AI and Research track (Library - 2nd floor) - Bangalore Internati onal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/near-real-time-feat ure-engineering-at-scale-for-machine-learning-use-cases-at-myntra-D8HzDhgf XsikWdcWLNifCC BEGIN:VALARM ACTION:display DESCRIPTION:Harmonising Art and AI: Crafting Jazzy and Juicy Video Snippet s through AI in AI and Research track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Unraveling the Identity Puzzle: Disambiguating Users across Social Media and Email Platforms DTSTART:20230811T094500Z DTEND:20230811T103000Z DTSTAMP:20260421T120126Z UID:session/X1sbyUwpQ4VbgQX8jKqnJk@hasgeek.com SEQUENCE:8 CATEGORIES:Confirmed CREATED:20230709T170709Z DESCRIPTION:Needl is an AI-assisted information hub that unifies data from diverse sources into one centralized platform\, which then allows for a v ariety of AI-assisted downstream workflows and collaboration across teams. Needl allows you to ingest\, organize and search your communications acr oss multiple platforms like gmail\, outlook\, telegram\, etc. Needl users come from all domains - they can be financial analysts\, legal experts or just general users.\n\nIn this talk\, we will discuss the problem to figu ring out if a Needl user is communicating with the same person across diff erent social media or email platforms. This is a difficult problem becau se a user can create multiple handles/accounts with radically different na mes on various platforms. We developed a solution which looks at account details\, the frequency of correspondence as well as the clusters of commu nication. The solution protects the privacy of Needl users while incorpor ating user feedback into the disambiguation process.\n\nThe system has bee n operational for about 2 years. For some users\, we obtained an F1-score of 0.78 on their medium and high priority contacts. In other cases\, we found that the solution had to be improved if multiple contacts of a Needl user had the same name (e.g. Anil) or if the name assigned by the Needl u ser to a given contact was different from their handle (e.g. email address ). We will present more details during the talk.\n\nThis is a joint talk by Sandeep Joshi and Vikram Srinivasan.\n\nSandeep Joshi currently does En gineering at Kognitos.com. This talk describes work done while working at Needl. \n\nVikram Srinivasan is the CEO and co-founder of Needl.ai (htt ps://www.needl.ai/). He is also an adjunct professor at IISc Bangalore\, where he teaches community detection algorithms among other things.\n\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20230828T114547Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/unraveling-the-iden tity-puzzle-disambiguating-users-across-social-media-and-email-platforms-X 1sbyUwpQ4VbgQX8jKqnJk BEGIN:VALARM ACTION:display DESCRIPTION:Unraveling the Identity Puzzle: Disambiguating Users across So cial Media and Email Platforms in Featured talks (Auditorium - first floor ) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Why LlamaIndex: Unlocking powers of LLMs DTSTART:20230811T094500Z DTEND:20230811T103000Z DTSTAMP:20260421T120126Z UID:session/ShPuPxoYWSX7zvnLVHuNd6@hasgeek.com SEQUENCE:15 CREATED:20230730T173850Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230908T114924Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Why LlamaIndex: Unlocking powers of LLMs in AI and Product tra ck\, Lightning talks and BOFs (Seminar halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:AI Society Panel: AI in India & Social Impact of AI DTSTART:20230811T094500Z DTEND:20230811T103000Z DTSTAMP:20260421T120126Z UID:session/3kWL57oZCWQksLt744nGt8@hasgeek.com SEQUENCE:10 CREATED:20230709T172647Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T063326Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:AI Society Panel: AI in India & Social Impact of AI in Industr ial track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Break DTSTART:20230811T103000Z DTEND:20230811T110000Z DTSTAMP:20260421T120126Z UID:session/BLzeULc6SABpNCiAwbQGwy@hasgeek.com SEQUENCE:2 CREATED:20230709T171252Z LAST-MODIFIED:20230709T171326Z LOCATION:Bangalore ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Break in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Space Models: Optimizing Store Space Allocation at Target DTSTART:20230811T110000Z DTEND:20230811T112500Z DTSTAMP:20260421T120126Z UID:session/2F2ZAXETdGWTE7QgKmQLAa@hasgeek.com SEQUENCE:11 CREATED:20230709T172734Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T023704Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Space Models: Optimizing Store Space Allocation at Target in I ndustrial track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:AI Career Panel: Stagewise HowTo\, New Roles et al DTSTART:20230811T110000Z DTEND:20230811T114500Z DTSTAMP:20260421T120126Z UID:session/U8fFYT7LGD6pnE34YJHuUa@hasgeek.com SEQUENCE:13 CREATED:20230709T173500Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230925T063255Z LOCATION:AI and Research track (Library - 2nd floor) - Bangalore Internati onal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:AI Career Panel: Stagewise HowTo\, New Roles et al in AI and R esearch track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Transforming Document Curation: LM and Vector Databases at Scale DTSTART:20230811T110000Z DTEND:20230811T114500Z DTSTAMP:20260421T120126Z UID:session/MPSUVW96xAHYvA6y9jqsea@hasgeek.com SEQUENCE:12 CREATED:20230709T172147Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230908T114856Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Transforming Document Curation: LM and Vector Databases at Sca le in AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sustainability: Design Considerations for Real-World ML-based Proc ess Model Training & IT/OT-to-SaaS Data Integration for Industrial Decarbo nization DTSTART:20230811T110000Z DTEND:20230811T114500Z DTSTAMP:20260421T120126Z UID:session/2GcANwvUsxqUJkziphjzYw@hasgeek.com SEQUENCE:5 CREATED:20230709T171303Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230830T123012Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Sustainability: Design Considerations for Real-World ML-based Process Model Training & IT/OT-to-SaaS Data Integration for Industrial Dec arbonization in Featured talks (Auditorium - first floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Multimodal AI: Representation and Reasoning on multimodal structur es DTSTART:20230811T114500Z DTEND:20230811T123000Z DTSTAMP:20260421T120126Z UID:session/LEhPKNUZjxTXhsCtRJQCMJ@hasgeek.com SEQUENCE:4 CREATED:20230730T174834Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230808T132953Z LOCATION:Industrial track (Board Room - 2nd floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Multimodal AI: Representation and Reasoning on multimodal stru ctures in Industrial track (Board Room - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Navigating the Credit Seas: A Unified Framework for Credit Risk Mo deling in CPG Industry DTSTART:20230811T114500Z DTEND:20230811T123000Z DTSTAMP:20260421T120126Z UID:session/2dpaJub7JdV4vWoJp7UgtC@hasgeek.com SEQUENCE:5 CREATED:20230709T173651Z GEO:12.9666826;77.6352903 LAST-MODIFIED:20230908T115203Z LOCATION:AI and Product track\, Lightning talks and BOFs (Seminar halls - 1st floor) - Bangalore International Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Navigating the Credit Seas: A Unified Framework for Credit Ris k Modeling in CPG Industry in AI and Product track\, Lightning talks and B OFs (Seminar halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Dynamic Control Match: Unleashing Adaptive Testing Power! DTSTART:20230811T114500Z DTEND:20230811T123000Z DTSTAMP:20260421T120126Z UID:session/DxsMjchRFbPNVnL4AAaVaC@hasgeek.com SEQUENCE:8 CATEGORIES:Confirmed CREATED:20230709T172327Z DESCRIPTION:A/B testing is a widely-used paradigm within marketing optimiz ation because it promises identification of causal effects\, because it is implemented out of the box in most messaging delivery software platforms\ , but mostly because it is held up as a "gold standard" for evaluating opt ions. This talk will explain why A/B tests are not a particularly sound me thod\, why businesses rarely choose better (adaptive) methods\, and outlin e the workings of a full-fledged testing approach that relies on reinforce ment learning to allow for thousands or even tens of thousands of simultan eous tests on overlapping samples of users. In particular\, we will focus on a method for disentangling causal effects of intermeshed tests under co nditions of continuous test adaptation\, using a matched-synthetic control group that adapts alongside the tests.\n GEO:12.9666826;77.6352903 LAST-MODIFIED:20240425T060039Z LOCATION:Featured talks (Auditorium - first floor) - Bangalore Internation al Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-08/schedule/validation-of-massi vely-parallel-adaptive-testing-using-dynamic-control-matching-DxsMjchRFbPN VnL4AAaVaC BEGIN:VALARM ACTION:display DESCRIPTION:Dynamic Control Match: Unleashing Adaptive Testing Power! in F eatured talks (Auditorium - first floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT END:VCALENDAR