Bhavishya Pandit

Emerging Stack for LLM-powered Product - Ideation [1/3] Sharing insights from Kuldeep Yadav's presentation on Large Language Models—Stack(LLM Stack). Dive into key learnings on building, deploying, and differentiating LLM-powered products in real-world applications. The secret sauce to successful LLM-powered products lies beyond the base model—it's about how you design, build, and continuously iterate your product lifecycle. From scoping real user needs & measurable business metrics, to prompt engineering and robust evaluation, every step defines differentiation. Remember, everyone accesses the same foundational models; your stack and methodology are the true competitive edges. Varun Aggarwal Shailendra Nath Jha Shubham Bansal Jitesh Bhasin Negrah S

13

How do we adapt large language models for structured, high precision tasks when traditional in context learning starts to fall apart. 🤔📊 In her talk “Beyond In-Context Learning: Rethinking Few-Shot Adaptation for Structured Tasks,” Dr. Sunita Sarawagi breaks down why ICL struggles with structured prediction and presents lighter, smarter alternatives from fine tuning to template constrained decoding. 🧠⚙️ Join us at ConfAI 2025, hosted by The Harish and Bina Shah School of AI and Computer Science at Plaksha University, Mohali from 28 to 30 November 2025. 🎙️📍 #AI #LLMs #MachineLearning #StructuredPrediction #ConfAI

15

🎥 𝗙𝗿𝗼𝗺 𝘁𝗵𝗲 𝗢𝗔𝗜 𝗔𝗿𝗰𝗵𝗶𝘃𝗲𝘀 If you're building or scaling large models, understanding how GPUs actually work together is non-negotiable. In this clip, Suraj Subramanian breaks down the mental models behind distributed training, clearly and practically 🔹 Data Parallelism – Splitting data across GPUs to scale training efficiently 🔹 Model Parallelism – Distributing model components across devices 🔹 Pipeline Parallelism – Structuring execution like an assembly line for inference Most importantly, he explains that these paradigms are orthogonal i.e. they don’t replace each other. They work together. This is the kind of foundational clarity that separates surface-level AI knowledge from production-ready ML engineering. 👉 Register for OAI 2026: https://lnkd.in/dh4mkdje 👉 Volunteer: https://lnkd.in/ewRTnz5U 👉 Join our Discord: https://lnkd.in/eDqBHamU #DistributedTraining #MachineLearning #LLMTraining #MLOps #OptimizedAIConference #ProductionAI

2

Chunking in NLP: Finding Structure in Sentences From translation to sentiment analysis, it all begins with identifying the right chunks. In this #TechBehindBhashini, Prof. Girish Nath Jha tells us how grouping words right helps machines translate, summarize, and extract meaning. #NLP #AITranslation #TechBehindBhashini #ChunkingInNLP Ministry of Electronics and Information Technology Narendra Modi Ashwini Vaishnaw Jitin Prasada S Krishnan Abhishek Singh Amitabh Nag Kavita Bhatia Digital India Programme Shailendra Pal Singh Ajay singh Rajawat Lureiliu Kamei Deepika Pathak Tarun Pandey IIT Bombay Indian Institute of Technology, Bombay Pushpak Bhattacharyya Vijay Kumar Krishna Kishore Asha Singh Jawaharlal Nehru University Jawaharlal Nehru Technological University AI4Bhārat IndiaAI India DST Abhay Karandikar Indian National Science Academy - INSA Indian Institute of Science (IISc) Council of Scientific and Industrial Research Indian Institute of Science Education and Research (IISER), Pune AU-KBC RESEARCH CENTRE Anna University Chennai International Institute of Information Technology Hyderabad (IIITH) Computation For Indian Language Technology (CFILT) Lab

50

1 Comment

Big things are happening with Snowflake, and at the recent Jaipur User Group meetup, the community came together to explore the latest AI innovations unveiled this year. Vinayak Mathur recaps how new features like Snowflake's Cortex Agents, AISQL, Openflow, and Horizon are already helping businesses solve real challenges in logistics, retail, customer support, and beyond. Read the full recap: https://lnkd.in/eVwEhCCN

33

2 Comments

Your 2026 AI roadmap probably skips the most critical piece: context. Model Context Protocol fixes that. It's not flashy, it's foundational. And when you see the CHAI insight, you'll understand why it matters more than the models themselves. Abhijit Joshi Sameer Karmarkar Chirag Barhate Ameya Varade Sandip Wane Anand Karwa Srikanth Sastry Anurag Rao Matt Jordan Jack Singer

6

Excited to announce that two new Hindi NLP datasets from ProcessVenue are now live on Hugging Face! 🇮🇳 These datasets use a double-blind annotation process and have high verifier accuracy. They provide reliable, ready-to-use Hindi data for the AI community. 𝗡𝗼𝘄 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲: 1. 𝗔𝗜-𝗔𝘀𝘀𝗶𝘀𝘁𝗲𝗱 𝗜𝗻𝘁𝗲𝗻𝘁 𝗖𝗹𝗮𝘀𝘀𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 – 5,000 natural Hindi prompts across 6 core intents 2. 𝗥𝗲𝘃𝗶𝗲𝘄𝗲𝗿 𝗦𝗲𝗻𝘁𝗶𝗺𝗲𝗻𝘁 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀 – 1,000 real Hindi movie reviews labeled for polarity. Perfect for training intent models, building sentiment systems, and powering research across the Hindi NLP ecosystem. 🔗 Find the Hugging Face links in the comments. #HindiNLP #AIDatasets #Dataset #DataAnnotation #ProcessVenue #HuggingFace #AIInnovation #MachineLearning #ArtificialIntelligence

33

2 Comments

From map testing to targeting pop-ups, Drizz is already being used across a range of product and growth teams. But these are just a few use cases among many. Yash Varyani explains. Watch the full video here: https://lnkd.in/gtEZJYTr

35

1 Comment

If you are serious about understanding how Large Language Models (LLMs) like GPT-3 and GPT-4 actually work under the hood, you need to stop scrolling and start watching this. What Dr. Raj Abhijit Dandekar (MIT PhD) has built at Vizuara is not a course. It is an open-access masterclass that dismantles the myth that LLMs are black boxes reserved for researchers inside OpenAI or Google. This playlist - titled "Building LLMs from Scratch" - has quietly become one of the most popular deep learning resources on the internet. Let the numbers speak: 👉 43 meticulously crafted lectures 👉 500,000+ views across the world 👉 Entire LLM pipeline built from scratch in Python This is not a surface-level explainer. This is from absolute zero to GPT-2 scale. You will: 👉 Build tokenizers (including Byte Pair Encoding) 👉 Create embedding layers 👉 Implement multi-head self-attention 👉 Code transformer blocks from scratch 👉 Train a 124-million parameter GPT-2 model 👉 Pretrain and fine-tune it 👉 Understand sampling methods like top-k and temperature scaling 👉 Work through full hands-on projects on classification and instruction fine-tuning Every lecture walks you through both the intuition and the implementation. Not with slides. With real code. And real understanding. There are no skipped steps. No hand-waving. Just intellectual honesty, teaching clarity, and world-class rigor. This playlist is already being used by ML researchers, startup teams, and university professors across the world. It has become a foundational reference for anyone building or understanding LLMs. If you want to break into AI or deepen your grasp of language models, this is the most valuable investment of your time. 📍Start here: https://lnkd.in/g36znuiu Built by Vizuara. Taught by Dr. Raj. Watched by the world. ____ 👉 Interested in learning AI/ML live from us? Check this out: https://lnkd.in/giEcTsMG 👉 Share with your connections if you found this useful 👉 Follow us for daily AI/ML updates and insights

14

Fractal, the Mumbai-based AI powerhouse, has unveiled Fathom-R1-14B, a 14B parameter open-source LLM that’s making waves with its math reasoning performance surpassing OpenAI's o1-mini and o3-mini, and nearly touching o4-mini—all at a post-training cost of just $499. Available on Hugging Face and licensed under MIT, Fathom-R1-14B scored 52.71% on AIME-25 and 35.26% on HMMT-25, with accuracy jumping to 76.7% and 56.7% respectively with inference-time boosts (cons@64). “This is just a tiny proof of what’s possible,” said Fractal CEO Srikanth Velamakanni, who shared that this effort is part of a larger proposal under the IndiaAI mission to build India’s first large reasoning model, including future plans for a 70B parameter model. Fathom-R1-14B was developed using SFT, curriculum learning, and model merging. A sibling variant, Fathom-R1-14B-RS, trained with reinforcement learning, showed similar results at a slightly higher cost of $967. This launch follows Fractal’s earlier AI healthcare initiative, Vaidya ai, and coincides with Sarvam AI’s release of Sarvam-M, a 24B parameter hybrid model for the IndiaAI mission. India’s open-source LLM race is heating up—and Fractal’s just thrown down the gauntlet.

2,713

54 Comments

From design to deployment, without the handoff hassle. AI Pundit Flow turns ideas into production-ready code, right inside your design environment. Faster iterations. Consistent output. Zero context loss. Here’s how it fits seamlessly into your workflow. Omkar Choudhari Sameer Ranabhor Shalvi Palande Ankita Kulkarni Kaushik Gokhale #DesignEngineering #AIPundit #TechInnovation #AIForDesign #WorkflowAutomation

53

1 Comment

Context is king. In this video, Aakarsh Kishore breaks down what sets MathCo’s AI approach apart and why prompting isn’t the only path to precision. Watch now to to hear how we’re making AI smarter by design. Ravit Jain #MathCo #NucliOS #ContextIsKing #AIWithContext #MathCoDifferentiator #OwnYourIntelligence

48

3 Comments

In this 20-minute talk, PolyMage Labs founder Uday Reddy Bondhugula talks about why and how one should build effective compilers for AI programming frameworks: https://lnkd.in/gWUQNUsF

79

A proud milestone in our journey with Databricks. Tredence is a launch partner for Databricks Lakebase.   We see Lakebase as the nervous system for real-time #ML and #AI, bringing transactional data, analytical insights, features, and models together with millisecond latency to power grounded, production-grade intelligence.   As a launch partner, we’re working hands-on with enterprises to modernize operational databases, build real-time, AI-native applications, and deploy agentic AI workflows on a single, governed foundation, where AI can reason, act, and adapt with trusted context. This partnership gives us a front-row seat to shape how the next generation of AI-native platforms is architected, not just imagined.   Excited to help define what real-time AI at enterprise scale looks like with Databricks. https://lnkd.in/gQK3TgDB Jason Y. | Amit Singh | Christine Gauthier | #Tredence #Databricks #BeyondPossible    

62

2 Comments

Struggling with the cost and complexity of building a dense knowledge graph for RAG? Partha Sarkar explains how to achieve high-recall retrieval with a simpler, more efficient GraphRAG approach.

7

In this video, Yashashvi breaks down why the industry is struggling to find true AI builders, people who understand data pipelines, model training, MLOps, cloud deployment, and real-world scaling. That’s exactly the gap BITS Pilani Digital aims to solve. Our AI Engineering & MLOPs programmes are designed with: • Hands-on industry projects that mirror production environments • Learning with tools used by top tech companies • Capstone projects that take AI all the way to deployment • Flexible online classes- Dual Mentorship program where you’re taught by both BITS faculty and industry experts. If your aim is to be a full-stack AI developer, explore the programme here: https://shorturl.at/qNPgw #OnlineEducation #DigitalLearning #BITSPilaniDigital #LearnDeeperRiseHigher Rajiv Tandon Aman Kishore Ramnarayan Banerjee Samit Dayal Smriti Sharma

14

Read Now: https://lnkd.in/dzHc_P8C Choosing between fine-tuning and contextual input isn’t easy—especially when accuracy, cost, and speed are all on the line. In this blog, our Senior Data Scientist, Alakh Sharma, breaks down when to use each strategy, sharing real-world examples and trade-offs to help you make the right decision for your use case. #AIModelTraining #FineTuning #GenerativeAI #DataScience #AIDevelopment

6

LLMs are powerful, until they melt your GPU budget and slow everything down. Today, Hrittik Roy and Aditya Soni will break down how they turned an unpredictable, resource-hungry LLM deployment into a fast, efficient, production-ready stack using CNCF tools. They’ll walk through: - Serving LLMs reliably with KServe + Kubeflow - Auto-scaling with Volcano and KEDA - Tracking performance + drift with Prometheus, Grafana & OpenTelemetry If you’re running LLMs in production... or trying to, this is a session worth joining. 🕒 2:30–3:00 PM EST 📍 Thomas Murphy Ballroom 4 | Level 5 #KubeCon #CloudNativeCon #vCluster

14

The future of data platforms lies in turning fragmented signals into a single, intelligent identity that adapts in real time. ruhail amin curates a forward-looking discussion with Satish (Dr) Mittal, Dilpreet Singh (CLMP™), Murthy Palli and Vinit Kapahi at #MartechAiConference. BW Businessworld chargeup ITC Hotels Limited Novartis AVIVA LIFE INSURANCE COMPANY INDIA LTD Here are the key takeaways : "True AI doesn’t just know what a customer wants, it understands why they want it, solving problems with empathy, not just marketing." - Vinit Kapahi, Chief Marketing Officer, Aviva India "Personalisation is powerful, but privacy is paramount, brands must know enough to delight, not to make customers feel watched." - Dilpreet Singh (CLMP™), Head of CRM, ITC Limited "AI should serve those who need it most, making every interaction practical and meaningful, especially for last-mile EV drivers." - Satish (Dr) Mittal, Co-Founder and Chief Digital Officer, chargeup "In a highly regulated industry like pharmaceuticals, digital engagement isn’t just a channel; it’s a responsibility. Success lies not in technology alone, but in marketers learning to ethically harness data to truly understand and educate both doctors and patients” - Murthy Palli, Group Head – MarTech Products, Novartis India #CustomerIntelligence #RealtimeCX #DigitalOrchestration #AIPowered #e4m

33