BEGIN:VCALENDAR VERSION:2.0 PRODID:-//HasGeek//NONSGML Funnel//EN DESCRIPTION:On the engineering and business implications of AI & ML X-WR-CALDESC:On the engineering and business implications of AI & ML NAME:The Fifth Elephant 2023 Winter X-WR-CALNAME:The Fifth Elephant 2023 Winter REFRESH-INTERVAL;VALUE=DURATION:PT12H SUMMARY:The Fifth Elephant 2023 Winter TIMEZONE-ID:Asia/Kolkata X-PUBLISHED-TTL:PT12H X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT SUMMARY:Check-in DTSTART:20231208T033000Z DTEND:20231208T040000Z DTSTAMP:20260421T120126Z UID:session/BhBAGyffhhs237Reb1Ss4R@hasgeek.com SEQUENCE:2 CREATED:20231122T051654Z LAST-MODIFIED:20231122T051706Z 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 Winter 2023 DTSTART:20231208T040000Z DTEND:20231208T041000Z DTSTAMP:20260421T120126Z UID:session/7wvBKDpSXRihZEdSwVzh1H@hasgeek.com SEQUENCE:0 CREATED:20231122T051738Z LAST-MODIFIED:20231122T051738Z LOCATION:Bangalore ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:Introduction to the Fifth Elephant Winter 2023 in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Chatting with Flipkart: How GenAI Revolutionised Online Shopping DTSTART:20231208T041000Z DTEND:20231208T045500Z DTSTAMP:20260421T120126Z UID:session/MoJD1WDDVne3T24wvmKxMm@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed,Data Science CREATED:20231122T052538Z DESCRIPTION:# Abstract\nIn India's offline markets\, shopping is always ab out having an interactive dialogue. Yet\, as e-commerce platforms evolved\ , dense UI interfaces clouded and removed conversational experience. With GenAI\, we at Flipakrt have revitalized this dialogue. We called it *Conve rsational Commerce*\n\nThis talk will dive into the technical nuances of t he GenAI implementation\, from semantic understanding\, to its seamless in tegration with Flipkart's vast product ecosystem. \n\n# Broad outline\n1. Touch upon the various use-cases where GenAI is used in Flipkart.\n1. Dive deep into how LLMs are used to **Understand User messages** better in *Co nversational Commerce*.\n 1. **Decoding Queries:** Use ChatGPT-like LLM models to interpret and act on user queries.\n 1. **Clarifying Ambigui ties:** Generate questions to gain clarity on ambiguous or vague inputs.\n 1. **Contextual Conversations:** Maintain continuity throughout dialog ues\, identifying when users switch topics.\n1. Deep dive into the importa nce of **Conversation Design** and how its used in-conjuction with LLMs t o provide a rich chat experience to the user.\n1. The challenges faced\, h ow they were overcome and the road forward.\n2. Finally\, our insights on LLMs\, how and when to use them\, and how local in-house models can augume nt large deployed models such as ChatGPT.\n\n# Target Audience\nData scien tists and Software Architects building Chat Assistants for large organisat ions with diverse use-cases.\n\n# Unique Value Proposition\nThe talk focus ses on building large scale **production** GenAI applications\, solving a broad set of problems in Flipkart. We will give a view of how and why a Ge nAI POC and a Production grade application are different.\n\n# Speakers\n* Anantharam Vanchiprakash [LinkedIn](https://www.linkedin.com/in/ananthara m-vanchi-prakash/)\n* Priyanka Nahata [LinkedIn](https://www.linkedin.com/ in/priyankanahata/)\n GEO:12.966671;77.6356638 LAST-MODIFIED:20240120T122421Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/chatting-with-flipk art-how-genai-revolutionised-online-shopping-MoJD1WDDVne3T24wvmKxMm BEGIN:VALARM ACTION:display DESCRIPTION:Chatting with Flipkart: How GenAI Revolutionised Online Shoppi ng in Data Science track (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Scaling Recommendations @ Meesho: Lessons from Twitter\, Instagram and Pinterest Ranking & Retrieval Strategies DTSTART:20231208T041000Z DTEND:20231208T045500Z DTSTAMP:20260421T120126Z UID:session/PFGLeHAzgHW2TQ6AqNS6VQ@hasgeek.com SEQUENCE:3 CATEGORIES:Confirmed,AI Engineering CREATED:20231122T052629Z DESCRIPTION:Abstract\nIn this session\, we'll deep-dive into Recommendatio n Systems\, with a special focus on two critical stages: Retrieval and Ran king. We'll begin by examining how different industries implement large-sc ale Retrieval Systems tailored to their specific needs. To illustrate our points\, we'll take a closer look at the Retrieval Systems used by Twitter and Instagram\, even though both are social networks\, their Recommender Systems are unique. We'll explore Twitter's follow suggestions and how Ins tagram recommends interactive media content to its users. We'll uncover th e strategies these companies use to create effective Retrieval Systems and identify common approaches.\n\nNext\, we'll shift our attention to Rankin g Systems\, where we'll explore how companies strategically order items to maximize conversions. We'll discuss the use of multi-objective ranking fu nctions\, a common approach in many companies. Additionally\, we'll also e xamine how other companies fine-tune their rankings using feedback loops t o align with multiple objectives. In this context\, we'll take a closer lo ok at the ranking systems of Instagram and Pinterest.\n\nThroughout this s ession\, we aim to provide insights into the world of Recommendation Syste ms and how in Meesho we have used some of these learnings to create high-q uality Recommender Systems.\n\nAbout the Speaker\nAbhishek Mungoli is a se asoned data scientist with over 7 years of experience\, holding a master's degree in Computer Science from IIIT-Hyderabad. He has worked with promin ent companies such as Walmart and currently serves as a Lead Data Scientis t at Meesho. Abhishek's expertise spans various domains of data science\, including supply chain\, pricing\, fraud analytics\, recommendation system s\, and advertising platforms. He is a thought leader in the field\, regul arly sharing his insights on platforms like LinkedIn and Medium\, as well as through his YouTube channel\, DataTrek. Abhishek has also delivered gue st lectures at prestigious institutions like IIT Madras\, Symbiosis Pune\, Jindal University \, and IIIT Sricity. He has build a modest following wi thin the Data Science community\, with around 19k+ followers on LinkedIn a nd approximately 3.7K+ subscribers on YouTube. Outside of his professional endeavors\, he's a fitness enthusiast and a devoted MMA fan. \n GEO:12.966671;77.6356638 LAST-MODIFIED:20231213T124710Z LOCATION:AI Engineering track (Library - 2nd floor) - Bangalore Internatio nal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/scaling-recommendat ions-meesho-lessons-from-twitter-instagram-and-pinterest-ranking-retrieval -strategies-PFGLeHAzgHW2TQ6AqNS6VQ BEGIN:VALARM ACTION:display DESCRIPTION:Scaling Recommendations @ Meesho: Lessons from Twitter\, Insta gram and Pinterest Ranking & Retrieval Strategies in AI Engineering track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Forecasting Architecture\, Metrics and Learnings @ Samsung Ads DTSTART:20231208T045500Z DTEND:20231208T054000Z DTSTAMP:20260421T120126Z UID:session/36ZxEBQ2oR1bo2teQZ7GHG@hasgeek.com SEQUENCE:12 CATEGORIES:Confirmed,Data Science CREATED:20231122T065208Z 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.\n\nSamsung Ads needs to sel l to advertisers in advance to show advertisements. The Ad opportunities d epend on user behaviour – users turning the TV on and going to specific screens. Having an automated way to forecast availability opportunities is critical to:\n• Ensure we do not over-commit to advertisers (monetary i mplications + hurts reputation)\n• Ensure we do not under-commit (opport unities are wasted\, potential revenue wasted)\n• Ensure users are able to self service the forecast process (tap into a larger market segment tha t wants self service\n\nOur goal is to predict how many impression events will be received for a specific campaign over the duration of the campaign . The campaigns are setup to target opportunities based on various criteri a including Location\, Time of day\, TV Model\, Type of ad opportunity and User identifiers.\nForecasting is a complex problem that typically involv es a single time series\, and to predict for one step ahead. We have many different time series patterns\, and multi-step forecasts with a long fore cast horizon (over 90 days). This requires the use of sequence to sequence models and modern techniques such as transformer architectures. We are bu ilding this solution using the state of the art Temporal Fusion Transforme r models. We will go over the different type of Ads such as Roadblocks\, A udience Take overs\, Rotationals and Video Ads and the factors affecting f orecasting.\n\nWe will go over the key challenges faced in coming up with a working model architecture\, such as erroneous ground truth data\, data availability & quality issues and data understanding gaps\, and our approa ches to deal with these challenges. We will go over the use of data sketch es and an OLAP DB like druid to get past data and use that and other featu res as inputs to a TFT model. For modeling external competition\, we will explain how we estimate price dependence of win rates using survival model s. We will also introduce the evaluation framework built to evaluate the f orecasting accuracy at three different phases - during development\, pre-r elease and post release. Earlier\, the analysis was done manually which ha d many challenges like lack of consistency\, delays\, lack of historical d ata etc. which were solved with the evaluation framework. GEO:12.966671;77.6356638 LAST-MODIFIED:20240120T122404Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/forecasting-archite cture-metrics-and-learnings-samsung-ads-36ZxEBQ2oR1bo2teQZ7GHG BEGIN:VALARM ACTION:display DESCRIPTION:Forecasting Architecture\, Metrics and Learnings @ Samsung Ads in Data Science track (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Video Highlights Generation DTSTART:20231208T045500Z DTEND:20231208T054000Z DTSTAMP:20260421T120126Z UID:session/7yTGqnjhY3vyqYT1dy4v5B@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed,AI Engineering CREATED:20231122T052723Z DESCRIPTION:### Abstract\nRoposo is a live video platform with over ~200 m illion end users with ~1000 live videos getting uploaded every day\, each lasting 15 minutes to 3 hours. In order to increase engagement and improve user experience\, we are trying to create a central video feed which will have assets that can easily consumed. This requires converting our events and creator led videos to shorter formats like trailers and short clips\, for which\, we process the videos with help of AI to get the most importa nt segments.\n\n### Broad Outline\nVideos can be very diverse as their con tent can vary from:\n - people having arguments\, dancing or singing (**Bi g Boss\, Glance being smart lock screen partner**)\n - just having convers ations like in an interview (**Creator Led Shows\, Exclusive content for R oposo**)\n - a fashion show where a supermodel just walk a runway (**Lakme Fashion Week\, Glance being a partner**).\n\nAs a solution to this\, we b ifurcated videos based on the density of speech happening in them and crea ted separate solutions for a speech-heavy and a visual-heavy video.\n\nFor a speech-heavy video\, we are use transcription to select the most import ant segments of a video while for a visual-heavy video\, we break videos i nto shots and generate visual descriptions of the shots to select the most important segments.\n\nWe are leveraging the following for our use-case:\ n - Faster [Whisper](https://arxiv.org/abs/2212.04356) using CTranslate2 f or audio transcription.\n - [BLIP](https://arxiv.org/abs/2201.12086) and [ Git](https://arxiv.org/abs/2205.14100) for Image Captioning.\n - [Vilt](ht tps://arxiv.org/pdf/2102.03334.pdf) for Visual Question Answering.\n - Col or Histograms for Shot boundary detection.\n - gpt3.5-turbo for text highl ights and summarisation.\n - [Sentence-BERT](https://arxiv.org/abs/1908.10 084) embedding and cosine similarity for retrival.\n - All the above model s optimised to run on a single T4 GPU using a custom dataloader for parall el processing.\n\nTo enhance the viewer experience\, we are post-processin g our short videos with AI-generated music\, custom transitions between sh ots\, animations\, stickers\, subtitles and a lot more.\n\nThe end-to-end processing runs at 5-10 mins for 30 min long video.\n\n### Impact\n - Incr eased content liquidity on our platform by 300%.\n - Increased average pla y duration (APD) on short videos by 44%.\n - Increased viewership for orig inal content by 23%.\n\n### Future Work\n - Introduction of multi-modality for describing segments.\n - Generalization across on more diverse videos . \n\n### Target Audience\nData Scientists and ML Engineers \n\n### Speake rs\n - [Shyam Choudhary](https://www.linkedin.com/in/schoudhary101/)\n - [ Akshat Gupta](https://www.linkedin.com/in/agupta28/) GEO:12.966671;77.6356638 LAST-MODIFIED:20231221T111007Z LOCATION:AI Engineering track (Library - 2nd floor) - Bangalore Internatio nal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/video-highlights-ge neration-7yTGqnjhY3vyqYT1dy4v5B BEGIN:VALARM ACTION:display DESCRIPTION:Video Highlights Generation in AI Engineering track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Break DTSTART:20231208T054000Z DTEND:20231208T060000Z DTSTAMP:20260421T120126Z UID:session/63QGk73QZLAzW9noJF3g9q@hasgeek.com SEQUENCE:1 CREATED:20231122T052755Z LAST-MODIFIED:20231122T052757Z 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:Shiksha Copilot- Empowering teachers to create best learning exper iences DTSTART:20231208T060000Z DTEND:20231208T064500Z DTSTAMP:20260421T120126Z UID:session/Tt53Jq1ya2XZzv5VJ2hH4V@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed,AI Engineering CREATED:20231122T052914Z DESCRIPTION:# Title\nShiksha copilot- Empowering teachers to create best l earning experiences \n\n## Overview\nShiksha copilot is a research project at Microsoft Research\, India. The project aims to improve learning outco mes and empower teachers to create comprehensive\, age-appropriate lesson plans combining the best available online resources\, including textbooks\ , videos\, classroom activities\, and student assessment tools. To help cu rate these resources\, the project team built a copilot—an AI-powered di gital assistant—centered around teachers’ specific needs.\n\n## Proble m\nTeachers are the backbone of any educational system. They are not just educators\; they are indispensable navigators\, mentors\, and leaders. Tea chers around the world face many challenges\, which vary from country to c ountry or even within a city or town. But some challenges are universal\, including time management\, classroom organization\, and creating effectiv e lesson plans\, personalizing content for henetrogenous student-base in t heir class and eventually improving the learning outcomes.\n\n## Solution\ nShiksha copilot is a copilot experience for teachers to generate personal ized and engaging learning experiences for their studnets in just a few mi nutes. Shiksha copilot addresses various societal and technical challenges in developing inclusive and accessible copilots such as improved multilin gual and multimodal interactions. Educational content is mainly multimodal \, including text\, images\, tables\, videos\, charts\, and interactive ma terials. Therefore\, for developing engaging learning experiences\, it is essential to build generative AI models which have unified multimodal capa bilities. Also\, these experiences are most impactful when delivered in na tive languages\, which requires improving the multilingual capabilities of generative AI models. \n\nThis copilot is being developed as part of Proj ect VeLLM (Universal Empowerment with Large Language Models) at Microsoft Research India. VeLLM’s goal is to make inclusive and accessible copilot s available to everyone by building a platform for developing population-s cale copilots. Inclusive copilots must address various real-world challeng es\, such as a multilingual user base\, varied skillsets\, limited devices and connectivity\, domain-specific understanding\, guardrails\, and safet y principles. Shiksha is the first copilot developed using the VeLLM platf orm. The VeLLM team is working with collaborators across diverse domains\, such as agriculture and healthcare\, to develop tailored domain-specific copilot experiences utilizing the platform and addressing associated resea rch problems. \n\n## Speaker\nTanuja Ganu\, Principal Research SDE Manage r\, Microsoft Research\, India\n GEO:12.966671;77.6356638 LAST-MODIFIED:20240120T122415Z LOCATION:AI Engineering track (Library - 2nd floor) - Bangalore Internatio nal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/shiksha-copilot-Tt5 3Jq1ya2XZzv5VJ2hH4V BEGIN:VALARM ACTION:display DESCRIPTION:Shiksha Copilot- Empowering teachers to create best learning e xperiences in AI Engineering track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:AI Maturity Continuum: A Three-Step Model to Understand Return On Investment (ROI) in AI DTSTART:20231208T060000Z DTEND:20231208T064500Z DTSTAMP:20260421T120126Z UID:session/M1TTQx6YUe1pUHyCcefLqu@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed,Data Science CREATED:20231122T052653Z DESCRIPTION:Building AI models/systems *doesn't* come cheap. Data\, Comput e Talent\, Time - it all comes at a high price in the AI world. Given the size of the investment\, be it a 0-1 startup or MNC\, it is important to u nderstand the ROI (Return On Investment) upfront.\n\nIn this talk\, we wil l look at "AI Maturity Continuum" - a framework that helps to conceive\, b uild & mature AI systems while keeping business outcomes in mind. To expla in the framework\, we will use the ROI curve in AI. \n\nWe will see how & why the *ROI curve in AI* is *very different* from the *ROI curve in tradi tional software/IT development*. This explains why the age-old corporate w isdom of "early wins" does not work in AI. Above all\, the framework guide s you on the kind of AI talent you must hire and how this changes with tim e. This directly impacts your AI hiring strategy!\n\n**Takeaways**\n1. A c oncrete yet simple framework to conceptualize "ROI" in AI projects\n2. Und erstand why ROI in traditional Software/IT behaves very differently for AI \n3. Understand how & why ROI for AI systems changes with time.\n4. Why wh ere you are on the ROI curve in AI has the greatest bearing on the kind of AI talent one needs to hire and How this changes with time\n5. Why contri butions margins in AI business are low and will never be anything like Saa S\n\n\nThanks\nAnuj \nhttps://www.linkedin.com/in/anujgupta-82/\n GEO:12.966671;77.6356638 LAST-MODIFIED:20240120T122410Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/ai-maturity-continu um-a-three-step-model-to-understand-return-on-investment-roi-in-ai-M1TTQx6 YUe1pUHyCcefLqu BEGIN:VALARM ACTION:display DESCRIPTION:AI Maturity Continuum: A Three-Step Model to Understand Return On Investment (ROI) in AI in Data Science track (Seminar Halls - 1st floo r) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:BoFs\, demos\, flash talks DTSTART:20231208T064500Z DTEND:20231208T073000Z DTSTAMP:20260421T120126Z UID:session/DBDNvt4fBouAg8PKuH7oWv@hasgeek.com SEQUENCE:1 CREATED:20231207T100938Z DESCRIPTION:Audience members can present demos or flash talks during this slot. GEO:12.966671;77.6356638 LAST-MODIFIED:20231207T100941Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com BEGIN:VALARM ACTION:display DESCRIPTION:BoFs\, demos\, flash talks in Data Science track (Seminar Hall s - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Conversational Search in Flipkart DTSTART:20231208T064500Z DTEND:20231208T073000Z DTSTAMP:20260421T120126Z UID:session/WstPUyfGj8robrAgADtAtV@hasgeek.com SEQUENCE:4 CATEGORIES:Confirmed,AI Engineering CREATED:20231122T053016Z DESCRIPTION:# Abstract\nWith Conversational Commerce becoming more popular \, the way people search for products on a Conversational (Chat) Interface is very different from the way they search on a traditional "Search Text Box" based interface. This talk is around the differences\, the challenges they pose and the solutionn to those challenges based on our learnings in Flipkart.\n\n# Broad Outline\nConversational Search is different from Tra ditional Search for the follwoing reasons.\n1. **Long\, descriptive and su bjective** queries - Eg. *“Show me best camera phones with good battery life”*. This is very different from traditional search queries which are keyword based. Eg. *"5g phones"*\n2. Suffer from **Articulation gap** - E g. *"Show me phones for daily use for my father*. Here\, we will need to u nderstand what features of a phone are required for daily use and for elde rs. These queries are not performed on traditional search methods and its not natural.\n3. Lots of **Comparison** queries - Eg. *"is vivo v15 bette r than oppo k10"*.\n4. Required **high relevance** - Since lesser results are are shown in a chat based search\, the results needs to be extremely r elevant.\n\nWe used LLMs for solving for articulation gap and query re-wri te. We also used different labelling techniques such as using LLMs to labe l the results rather than relying on metrics such as CTR to get high quali ty labelled results to train our models.\n\nWe will cover all of the above in the talk.\n\n# Target Audience\nSoftware Engineers and Data scients wh o are working on AI solutions for Search\, Comparison and Recommendation problems.\n\n# Speakers\nAnantharam Vanchiprakash - [LinkedIn](https://www .linkedin.com/in/anantharam-vanchi-prakash/)\nSajal Gupta - [LinkedIn] (ht tps://www.linkedin.com/in/sajalgupta0102/)\nAmey Patil - [LinkedIn] (https ://www.linkedin.com/in/amey-patil-692245128/)\nShreyas Shetty - [LinkedIn] (https://www.linkedin.com/in/shreyasshetty/)\n\n GEO:12.966671;77.6356638 LAST-MODIFIED:20231213T125930Z LOCATION:AI Engineering track (Library - 2nd floor) - Bangalore Internatio nal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/conversational-sear ch-in-flipkart-WstPUyfGj8robrAgADtAtV BEGIN:VALARM ACTION:display DESCRIPTION:Conversational Search in Flipkart in AI Engineering track (Lib rary - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Lunch DTSTART:20231208T073000Z DTEND:20231208T083000Z DTSTAMP:20260421T120126Z UID:session/6x1cVHLrYmpdW3sD1FbL9q@hasgeek.com SEQUENCE:0 CREATED:20231122T053213Z LAST-MODIFIED:20231122T053213Z 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:Economics of AI DTSTART:20231208T083000Z DTEND:20231208T091500Z DTSTAMP:20260421T120126Z UID:session/85GpKQYEP52UPdDaarxTL6@hasgeek.com SEQUENCE:2 CATEGORIES:Confirmed,BoF/Panel Discussion CREATED:20231205T083245Z DESCRIPTION:It is widely believed that the next few unicorns and decacorns will come from AI. This should hardly be surprising - AI is one of those once-in-a-lifetime technological advances that can change the life of ever y human on earth. Its impact on mankind will be no less than the impact of industrialization or electricity. \n\nBut building AI is one thing and bu ilding the business of AI is completely another beast. Building successful companies requires a solid business model and not just game-changing tech nology. The unit economics must work out well. \n\nIn this talk\, we will look at why the business of AI so far is super tricky. We will closely und erstand: \n1) How n why AI is a very expensive technology\n2) We will look at its various components from a cost lens\n3) We will understand why\, u nlike SaaS\, cost margins for AI business is expected to be super low\n\nW e hope that this talk will help founders\, leaders\, AI teams\, look at AI from a PnL lens and develop a more pragmatic view of the business of AI s o as to enable the broader AI community to make much more informed decisio ns.\n\nThanks\nAnuj\nhttps://www.linkedin.com/in/anujgupta-82/ GEO:12.966671;77.6356638 LAST-MODIFIED:20231215T075506Z LOCATION:AI Engineering track (Library - 2nd floor) - Bangalore Internatio nal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/economics-of-ai-85G pKQYEP52UPdDaarxTL6 BEGIN:VALARM ACTION:display DESCRIPTION:Economics of AI in AI Engineering track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:How Differential Privacy Changed The World DTSTART:20231208T083000Z DTEND:20231208T091500Z DTSTAMP:20260421T120126Z UID:session/RMArQPy24tUpvE2MzxcrWN@hasgeek.com SEQUENCE:11 CATEGORIES:Confirmed,Data Science CREATED:20231122T052837Z DESCRIPTION:## How Differential Privacy Changed The World\n##### and how e ngineering it into your pipelines can lead you to comply with legal requir ements and meet consumer wants and needs.\n\n+ We finally have a legal fra mework in India\, The Digital Personal Data Protection Act 2023\, which pr esents GDPR like requirements for Data Governance and Personally Identifia ble Information Protection.\n\n+ Our agenda with this talk is to look into Differential Privacy\, a game changing approach to robust and mathematica lly rigorous data privacy preservation. We pitch it as a practical solutio n\, and at the same time\, look into associated risks and ways to cope\, w ith the privacy-utility being the basis of the tradeoff.\n\n+ As a case st udy\, we look at how Wikipedia used differential privacy to release aggreg ate statistics in a privacy preserved manner.\n\n+ We will discuss how wit h the help of libraries like PyDP\, OpenDP or PyTorch Opacus and Tensorflo w Privacy\, one can work towards incorporating sturdy privacy practices in to datasets and pipelines to serve this need. as asked for by consumers\, as well as comply with legal requirements that in the coming couple of yea rs\, even startups will be asked to in a stringent manner.\n GEO:12.966671;77.6356638 LAST-MODIFIED:20231210T101150Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/how-differential-pr ivacy-changed-the-world-RMArQPy24tUpvE2MzxcrWN BEGIN:VALARM ACTION:display DESCRIPTION:How Differential Privacy Changed The World in Data Science tra ck (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Automated Genre Based Music Morphing using AI DTSTART:20231208T091500Z DTEND:20231208T093000Z DTSTAMP:20260421T120126Z UID:session/Rh1GAzt3urHvPVk1PCZdxp@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed,Demo CREATED:20231122T053318Z DESCRIPTION:Abstract\n\nIn the realm of audio processing and music product ion\, automated genre-based audio morphing is an emerging field that merge s the creative boundaries of different music genres through the power of m achine learning and natural language processing (NLP). This innovative app roach leverages textual prompts to drive the transformation of audio conte nt\, transcending traditional genre constraints and enabling new forms of musical expression. This can then be used as an isolated content for perso nalized recommendation\n\n\nApproach\n\nLeveraging state of the art models \n\n1. MusicGen (by Meta) - To drive the music conditioning based on pro mpts\n2. LP-MusicCaps: LLM-Based Pseudo Music Captioning - Retrieve taggin g based on music attributes\n\nWe finetune them on our dataset\, which we created weakly using ChatGPT and in house annotations. We also used 3P dat asets like WavCaps to augment our dataset iwth real world examples. We ach ieve audio morphing in ~30 seconds for a 5 min video.\n\n\nDeployment at s cale\n\nWe are deploying at scale in near future\, but this will directly impact the liquidity\, we have atleast 5-6 variants for the same audio (wh ich would mean 5-6x liquidity) which then we can use in downstream tasks ( like personalised recommendation\, etc)\n\n\nKey Points to be discussed in the talk- \n\n1. End to End pipeline for music morphing\n2. Finetuning on our dataset\n3. Challenges faced in terms of deployment (multi GPU) + red ucing latencies.\n4. How this can be used as a personalised content for re commendation\n\nSample Example\nmusic input https://drive.google.com/file/ d/1iEN0uML7FMxbKbO3uJckFnJkMtB12aH5/view?usp=sharing\nTheme - beats funk 8 0s music\nOutput - https://drive.google.com/file/d/13HX1Rrka1AgSCXdBsd0e_k Ci2XBffINi/view?usp=sharing\n\n\n GEO:12.966671;77.6356638 LAST-MODIFIED:20231221T060511Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/automated-genre-bas ed-music-morphing-using-ai-Rh1GAzt3urHvPVk1PCZdxp BEGIN:VALARM ACTION:display DESCRIPTION:Automated Genre Based Music Morphing using AI in Data Science track (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Revolutionizing Radiology Reports: Leveraging AI and LLMs for Impr oved Quality and Actionability DTSTART:20231208T091500Z DTEND:20231208T100000Z DTSTAMP:20260421T120126Z UID:session/V7Vh3hZdTygZKvCjypGWJN@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed,AI Engineering CREATED:20231122T053404Z DESCRIPTION:### Overview\nIn the era of data-driven healthcare\, the quali ty and actionability of healthcare reports are critical. With improved acc ess to diagnostics\, radiology plays a pivotal role in adminstering care. \n\n\n### Problem\nFor patients who've had a radiology scan for their cond ition\, the radiology reports serve as the basis for treatment plans\, an d hence their clarity\, accuracy\, and detail are of utmost importance. Ho wever\, the reports can often be riddled with jargon\, ambiguities\, or in sufficient detail\, leading to potential misunderstandings or misdiagnoses . \n\n### Solution\n\nIntegrating LLMs into the radiology workflow address es these issues\, aiding radiologists in crafting more precise\, understan dable\, and actionable reports. This ultimately contributes to better pati ent outcomes and more efficient healthcare systems.\n\nThis talk will eluc idate the transformative role of Large Language Models (LLMs) in improving the quality and actionability of radiology reports. The talk is based on the speakers' work at their organization\, 5C Network - India's largest AI -driven Clinical Delivery Network. \n\nWe'll explore real-world case studi es\, examine the technical architecture\, and discuss the benefits and cha llenges that come with the integration of LLMs in radiology workflows.\n\n ### Agenda\n* Introduction \n * The crucial role of radiology reports i n healthcare\n * Challenges in traditional radiology reporting\n* Why L arge Language Models? \n * Case for LLMs in radiology\n* Technical Arch itecture\n * End-to-end workflow of integrating LLMs in radiology\n * API interfaces\, data preprocessing\n * model fine-tuning\n * self -hosting \n* Real-life success stories/case studies \n* Metrics to evaluat e and improve report quality and actionability\n* data privacy\n* ROI\, sc alability (depending on time)\n\n\nBy the end of the presentation\, attend ees will have a comprehensive understanding of the role of LLMs in enhanci ng radiology reports. They will be equipped with actionable insights and b est practices for implementing these models in their own systems.\n\n### S peakers\nKalyan Sivasailam - Co-founder and CEO\, 5C Network\nBargava Subr amanian - Chief Product and Data Officer\, 5C Network\n\n\n\n\n\n\n\n GEO:12.966671;77.6356638 LAST-MODIFIED:20240120T122359Z LOCATION:AI Engineering track (Library - 2nd floor) - Bangalore Internatio nal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/revolutionizing-rad iology-reports-leveraging-ai-and-llms-for-improved-quality-and-actionabili ty-V7Vh3hZdTygZKvCjypGWJN BEGIN:VALARM ACTION:display DESCRIPTION:Revolutionizing Radiology Reports: Leveraging AI and LLMs for Improved Quality and Actionability in AI Engineering track (Library - 2nd floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Accelerating Generative AI initiatives with AMD SharkStudio Techno logy DTSTART:20231208T093000Z DTEND:20231208T094500Z DTSTAMP:20260421T120126Z UID:session/P2DFL8qx2WHNja64phBbo9@hasgeek.com SEQUENCE:7 CATEGORIES:Confirmed,Demo CREATED:20231122T053429Z DESCRIPTION:Discover AMD SharkStudio\, a tool to simplify Generative AI in itiatives. With an easy-to-use interface\, SharkStudio empowers users to i nteract effortlessly with open source LLMs and Stable Diffusion Models\, d elivering enhanced performance on AMD hardware. SharkStudio provides a low friction deployment from laptops (Ryzen) to datacenters (EPYC CPU\, and M I210/MI250 GPU). \n\nFind out more about SharkStudio and the technology th at it delivers in this talk. GEO:12.966671;77.6356638 LAST-MODIFIED:20231221T110711Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/accelerating-genera tive-ai-initiatives-with-amd-sharkstudio-technology-P2DFL8qx2WHNja64phBbo9 BEGIN:VALARM ACTION:display DESCRIPTION:Accelerating Generative AI initiatives with AMD SharkStudio Te chnology in Data Science track (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Building LLM Apps up to 20x Faster using Microservices Architectur e DTSTART:20231208T094500Z DTEND:20231208T100000Z DTSTAMP:20260421T120126Z UID:session/Ryu5XBMwBGyQDG7W3fRYgQ@hasgeek.com SEQUENCE:10 CATEGORIES:Confirmed,Demo CREATED:20231122T053456Z DESCRIPTION:## Overview\n\nMicroservices Architecture has transformed API and Frontend development\, extending its impact to billions of users. Now\ , with the advent of generative AI and Large Language Models (LLMs)\, ther e's an unprecedented opportunity to disrupt industries. However\, developi ng reliable and safe LLM applications poses significant challenges for dev elopers. This talk explores how Microservices Architecture can address the se challenges and accelerate LLM app development.\n\n## Problem\n\nDevelop ers face multiple challenges when building scalable and safe LLM applicati ons:\n\n1. LLM Selection\n2. LLM Interoperability\n3. Prompt Reusability\n 4. Prompt Collaboration\n5. Prompt Testing and Evaluation\n6. Prompt Refin ing and Fine-Tuning Cost\n7. Reproducibility\n8. Guardrailing\n9. Outdated Knowledge of LLMs\n\n## Solution\n\nThis talk proposes breaking down the monolithic approach into a layered architecture inspired by Microservices. This approach addresses distinct concerns for app developers\, prompt dev elopers\, and data scientists. Key components of Microservices Architectur e and open-source frameworks for Prompt Management\, including versioning\ , refining\, accuracy\, backtesting\, and training coupled with MicroLLM s ervice\, will be discussed.\n\nThis talk is based on the speaker work at [ Sugarcane AI](https://sugarcaneai.dev/) - npm like ecosystem for prompts.\ n\n## Agenda\n\n- **Introduction**\n- **Challenges with Monolithic LLM De velopment**\n- **Microservices Architecture in LLM**\n - Components of Mi croservices Architecture\n- **How Microservices Architecture accelerates d evelopment by up to 20x**\n -. LLM Selection\n - LLM Interoperability\n - Reusability\n - Cost Challenges\n - Reproducibility\n - Guardrailing \n - Addressing Outdated Knowledge of LLMs\n - Migrating to Microservice s Architecture in LLM Apps\n - Time and Cost of Development and Maintenan ce\n- **Cherry on the Top**\n - AI Safety\n - Knowledge Upgradation\n - Real-world Examples and Challenges (depends on time)\n\n## Speakers\n\n- *Ankur Agarwal - Founder\, [Sugarcane AI](https://sugarcaneai.dev/)*\, [Li nkedIn](https://www.linkedin.com/in/ankuragarwalmnnit)\n## Conclusion\n\nA ttendees will gain a comprehensive understanding of leveraging Microservic es Architecture to build LLM applications up to 20x faster. Real-world exa mples and insights from Sugarcane AI and Toast will provide practical know ledge for navigating the challenges of LLM-based app development.\n GEO:12.966671;77.6356638 LAST-MODIFIED:20240207T090417Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/building-llm-apps-u p-to-20x-faster-using-microservices-architecture-Ryu5XBMwBGyQDG7W3fRYgQ BEGIN:VALARM ACTION:display DESCRIPTION:Building LLM Apps up to 20x Faster using Microservices Archite cture in Data Science track (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Near Real time Feature Engineering for Machine Learning use cases at Myntra Scale DTSTART:20231208T100000Z DTEND:20231208T104500Z DTSTAMP:20260421T120126Z UID:session/8CuaicQGr5HrhySK16L734@hasgeek.com SEQUENCE:3 CATEGORIES:Confirmed,AI Engineering CREATED:20231122T053717Z DESCRIPTION:# Problem\n\nMyntra is one of the leading fashion e-commerce c ompanies in India. Myntra delivers best-in-class shopping experience by le veraging many advanced machine learning models\, deployed for online or re al-time inference. The online inference requires streams of data to be pro cessed\, machine learning features to be computed\, stored and served in ( near) real-time\, at Myntra scale.\n\nThe features can be hand crafted or generated (e.g. user\, product\, style\, widget\, image embeddings). And m ajority of the features require stateful stream processing with complex co mputation\, in (near) real-time\, at very high throughputs (millions of rp m) and low latency. This requires scalable\, resilient data engineering sy stems with stateful stream processing capabilities and feature stores.\n\n \n# Solution\n\nMyntra Data Engineering team designed and built Quicksilve r\, a real time data ingestion and stateful stream processing platform. It is part of the overall Myntra Data Platform. The Quicksilver platform ing ests millions of events every minute\, computes the machine learning featu res in (near) real time and makes them available to machine learning model s for online inference.\n\n\n# Outline of the talk\n\nOnline ML use cases at Myntra\nLife cycle of an online ML model\, including feature engineerin g\nChallenges of realtime feature engineering at scale\nFunctional and non -functional requirements\nArchitecture of the QuickSilver platform\, desig n principles and tech choices\nIntegration with Machine learning platform\ nBest practices and learnings\n GEO:12.966671;77.6356638 LAST-MODIFIED:20231215T080938Z LOCATION:AI Engineering track (Library - 2nd floor) - Bangalore Internatio nal Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/near-real-time-feat ure-engineering-for-machine-learning-use-cases-at-myntra-scale-8CuaicQGr5H rhySK16L734 BEGIN:VALARM ACTION:display DESCRIPTION:Near Real time Feature Engineering for Machine Learning use ca ses at Myntra Scale in AI Engineering track (Library - 2nd floor) in 5 min utes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:AI Powered Advertising DTSTART:20231208T100000Z DTEND:20231208T101500Z DTSTAMP:20260421T120126Z UID:session/L5bN5LzQysSsWry43KDEgQ@hasgeek.com SEQUENCE:4 CATEGORIES:Confirmed,Lightning Talk CREATED:20231122T053554Z DESCRIPTION:# Abstract (Problem)\nScaling digital ads heavily relies on te sting new ads. Many brands are unable to scale even though they have budge ts due to lack of ads to test. The biggest challenge for any advertiser is to come up with new Ad ideas\, design those Ads and track them at a fast pace so that the ad costs can be reduced.\n\n# Broad Outline (Solution)\nT here are several steps an advertiser needs to take to overcome the Ad crea tion challenge\n1. Generating new ideas - Advertisers need to come up with new ideas ever so often. This creates a roadblock in generating new ads. We use LLMs to generate new ideas for a brand that have a high chance of c onverting the audience into customers.\n2. Creating Ads - Based on the new ideas - generating high converting copy and creatives is a challenging an d time consuming task. We help generate new high converting Ad copy and cr eatives using a range of AI tools to speed up the process.\n3. Tracking Ad s - While we generate new Ads\, it is important to track every Ad and meas ure its performance at a much deeper level so that we can drive actionable insights for our next iteration of Ad Creatives. Our Deep AI Attribution model helps Advertisers with a dashboard on providing deeper insights abou t their Ad’s performance.\n4. Audience Suggestions - Selecting the right target audience is a crucial part in ensuring the performance of ads. How ever\, most targeting options are hidden behind API calls which makes adve rtisers unaware of untapped markets. Using the power of LLMs\, we provide tailored targeting options based on the offer so that the advertisers find the message to market fit.\n\n\n# Target Audience\nAdvertisers and Data S cientists who are working in the field of Marketing.\n\n# Slides\nhttps:// docs.google.com/presentation/d/1tO9q5sZxeDKrNpp7LXEfCZxyWeC8Nib4N2qfV4S3LW M/edit?usp=sharing\n\n# Speakers\nSantosh GSK - https://www.linkedin.com/i n/santoshgsk\n GEO:12.966671;77.6356638 LAST-MODIFIED:20231221T111757Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/ai-powered-advertis ing-L5bN5LzQysSsWry43KDEgQ BEGIN:VALARM ACTION:display DESCRIPTION:AI Powered Advertising in Data Science track (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Automating Knowledge Extraction for Training Content Generation wi th LLMs @ SquadStack DTSTART:20231208T101500Z DTEND:20231208T103000Z DTSTAMP:20260421T120126Z UID:session/PvP8AgjAq9GXHshny6msRP@hasgeek.com SEQUENCE:6 CATEGORIES:Confirmed,Lightning Talk CREATED:20231122T053614Z DESCRIPTION:### Abstract\nIn traditional BPOs\, training sales representat ives is challenging\, primarily due to limited visibility into the univers e of knowledge that the representatives have to stay on top of. This limit ation often results in suboptimal sales interactions with customers. Furth ermore\, the creation of thorough and timely training materials is a time- consuming endeavor\, even before the actual training can commence. Consequ ently\, this extended preparatory phase delays sales representatives from promptly engaging in customer calls on behalf of the business\, resulting in missed opportunities.\n\nAs a solution to this problem\, [SquadStack](h ttps://www.squadstack.com/) is in the process of creating an in-house tool \, we call **Intel Assist**\, to streamline the extraction of content\, fo r faster and accurate generation and querying of training material for a v ariety of businesses at a very low cost. This as a result\, is empowering our sales representatives to access and utilize this information seamless ly during customer calls.\n\nAt its core\, we aim for source-agnostic know ledge integration\, gathering information from a wide range of sources to enhance documentation and streamline the process of querying this knowledg e. While we have initially applied this approach to enhance our internal b usiness processes at SquadStack\, it can also benefit other enterprises se eking to bolster their own knowledge management capabilities.\n\n### Featu res\n**Step 1 - Knowledge Onboarding** \n- Extracting asked Questions\, Qu estions & Answers (QnAs)\, and Summaries from a variety of raw data source s\, including Call Recordings\, Word Documents\, PDFs\, Spreadsheets\, and Presentations.\n- Our comprehensive logging of each question and objectio n during calls has resulted in a valuable data repository\, opening the do or to a multitude of analytical insights.\n- By identifying frequently ask ed questions and objections\, we equip our representatives with the means to improve their preparedness and performance.\n \n**Step 2 - Knowledge S ufficiency**\n- Automatically generate a report assessing the alignment of the extracted knowledge with periodic changes and new requirements.\n- We can precisely discern which of the current questions being posed can be a ddressed using our existing knowledge base and which ones cannot.\n- Subse quently\, we can present these specific questions to our customers\, gathe r their responses\, and incorporate the newfound insights back into our kn owledge repository.\n\n**Step 3 - Knowledge Search**\n- Swiftly interrogat e an extensive knowledge database spanning multiple businesses.\n- This pr oves highly effective when sales representatives require real-time access to information that transcends conventional keyword matching.\n\n### Techn ology\n- [In-House ASR](https://aws.amazon.com/marketplace/pp/prodview-syf o3zgpq63bk?sr=0-1&ref_=beagle&applicationId=AWSMPContessa) for Transcripti on (Presented at Fifth Elephant Monsoon ‘23)\n- LLMs\n- [Sentence Transf ormers](https://huggingface.co/sentence-transformers) for generating embed dings\n\n### Impact\n- Reduction in time required to acquire knowledge\n- Reduction in the expenses to qcquire knowledge\n- Accelerated onboarding f or new customers\, leading to quicker Return on Investment (ROI) delivery for our clients.\n\n### Flow of Presentation\n- SquadStack’s Introductio n (1-2 mins)\n\n- Product Introduction (5-8 mins)\n - Industrial Problems (3-5 mins)\n - Overview of our solution - Intel Assist (2-3 mins)\n - I ntroduce ASR as previously showcased in Fifth Elephant Monsoon ‘23 (2-3 mins)\n- Features (7-12 mins)\n - Knowledge Onboarding (3-5 mins)\n - Kn owledge Sufficiency (3-5 mins)\n - Knowledge Search (1-2 mins)\n- Applica tion’s Lifecycle (3-5 mins)\n - What is the direct output of Intel Assi st? (Streamlit App)\n - How do we provide training?\n - Impact (1-2 mins )\n - Time Saved\n - How can people make use of a tool like this\n- Audi ence QnA\n\n### Speakers\n[Sumit Saha](https://in.linkedin.com/in/linksumi tsaha)\n GEO:12.966671;77.6356638 LAST-MODIFIED:20231221T110345Z LOCATION:Data Science track (Seminar Halls - 1st floor) - Bangalore Intern ational Centre (BIC)\nBengaluru\nIN ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com URL:https://hasgeek.com/fifthelephant/2023-12/schedule/automating-knowledg e-extraction-for-training-content-generation-with-llms-squadstack-PvP8AgjA q9GXHshny6msRP BEGIN:VALARM ACTION:display DESCRIPTION:Automating Knowledge Extraction for Training Content Generatio n with LLMs @ SquadStack in Data Science track (Seminar Halls - 1st floor) in 5 minutes TRIGGER:-PT5M END:VALARM END:VEVENT END:VCALENDAR