Nitin Bhandari

🚀 Just wrapped up two exciting AI projects as part of my Generative AI & ML journey — and I'm genuinely proud of what I built. 🛍️ Real-Time Retail Feedback Intelligence (Capstone — Generative AI) I built an end-to-end GenAI system for a retail brand (ChicStyle) that analyses thousands of customer reviews in real time. Using Zero-Shot, Few-Shot, and Chain-of-Thought prompting, the system: ✅ Classifies feedback as positive, negative, or neutral ✅ Identifies the product or service being reviewed ✅ Flags urgent issues automatically ✅ Generates personalised customer replies The biggest insight? Choosing the RIGHT prompting strategy matters as much as the model itself. Chain-of-Thought was a game-changer for nuanced sentiment cases. 🛒 Amazon Product Recommendation System (Elective — Recommendation Systems) Using real Amazon ratings data, I built a collaborative filtering engine powered by Matrix Factorisation (SVD) that cuts through information overload and surfaces genuinely relevant products for each user. What I took away from both projects: → AI is only as good as the problem framing behind it → Real-world data is messy — EDA and preprocessing are underrated superpowers → The gap between a working model and a useful product is closed by design thinking If you're exploring Generative AI, NLP, or recommendation systems — I'd love to connect and share notes! 🙌 🔗 Full portfolio: https://lnkd.in/dfX7b7zi #MIT #GenerativeAI #PromptEngineering #MachineLearning #NLP #SentimentAnalysis #RecommendationSystems #AIProjects #DataScience #Capstone #GreatLearning #AIForBusiness #OpenToWork #TechPortfolio

14

The Nvidia-only era of AI is taking its first real hit? Another insightful piece from Dylan Patel and SemiAnalysis, and one that gets an important point right: building a full AI accelerator stack where silicon, systems, and software are designed together really matters. The article still treats dense compute as the natural center of gravity. But the field isn’t converging, it’s diversifying. As AI models and workloads evolve, we are moving from “brute-force” dense computing toward approaches that are more sparse, more efficient, and far better for cost and energy use. It’s also a wake-up call for the entities signing million- and billion-dollar GPU deals without questioning whether those architectures actually fit their future workloads. The future looks broader, not just bigger. https://lnkd.in/eqP7vpAC

12

We are back with RAG workshop 🔥 After the incredible response to our MCP, Agentic AI, and Vibe Coding sessions, Arvind Narayanamurthy and I are kicking off a brand-new workshop series on one of the most requested and practical topics- Retrieval-Augmented Generation (RAG). The first installment of this series is specifically designed for beginner to intermediate learners, offering a comprehensive introduction to the fundamental concepts of RAG. 🔍 Whether you're building production-grade AI apps or just starting to explore how LLMs can go beyond their training data - RAG is a must-have skill. This isn’t your usual high-level webinar. We’ll take you through real implementation, live guidance, and building from scratch You’ll: ✅ Learn how RAG improves LLM performance with fresh, contextual data ✅ Understand vector DBs, indexing, and retrieval techniques ✅ Get Hands-on: Build and deploy a working RAG pipeline using LangChain + Pinecone ✅ See no-code and low-code demos for faster experimentation ✅ Explore real-world use cases and performance tips 📅 Reserve your seat, roll up your sleeves, and let's get building. 👉 https://lnkd.in/gS_B7_7d

151

15 Comments

What Oppenheimer’s Gita Moment Teaches Us About AI Ethics and Leadership https://lnkd.in/gM6ZAmx6 Global AI Gurgaon Global AI Community Gartner IndiaAI AI For Companies AI For Executives AI For Leaders #AI

12

Had a great experience attending the Enterprise AI Summit in San Jose, CA 🇺🇸 Got the opportunity to interact with industry leaders like Gene Kim, Steve, Kent Beck, and professionals from organizations such as Amazon and Netflix. The conversations around Agentic AI workflows and their impact on how products are built and scaled were particularly thought-provoking. One key takeaway: We are moving beyond traditional automation to systems that can reason, act, and continuously evolve — fundamentally reshaping product thinking and execution. A special highlight was meeting Gene Kim in person — a quick interaction, but definitely a memorable one. Looking forward to applying these learnings to build more intelligent, adaptive, and outcome-driven products. #EnterpriseAI #AgenticAI #ProductManagement #AI #Innovation #FutureOfWork

140

In my interactions over the last few days at the India AI Impact Summit, it seems like there is a great potential in integrating generative AI (RAG, agents, contextual recommendations, multi-modal / multi-lingual models etc) with ECHO's community-based learning model to further bridge the knowledge divide - however we must work to empower front line workers instead of replacing them, while making communities more resilient towards future pandemics and other crises. Impact Ignited- AI Powered Ideas in Action at InvokED 5.0 brought together insightful ideas from a diverse group, where I shared how Project ECHO is developing foundational technology to build and grow trusted human networks — amplifying the human impact of the 100 million+ front line workers globally A.I. is not a silver bullet that will magically solve challenges of inequity — it however can play a big role in democratizing access and empowering FLWs and eliminate barriers of region, language, and tech-literacy. This is the core of what our work at Project ECHO is all about.

166

3 Comments

Scaling LLMs for the Enterprise: Why Contextual Density Matters More Than Scale. We’ve reached a point where increasing model parameters yields diminishing returns if the underlying retrieval architecture is flat. I’m looking forward to speaking at Sri Ramakrishna Institute of Technology (SRIT) on 29 December about GraphRAG and the shift toward structured knowledge representations. We will move beyond simple semantic similarity to discuss: ➔ Relational Retrieval: Navigating high-order dependencies that vector-only RAG misses. ➔ Global Contextualization: Solving the "needle in a haystack" problem by summarizing entire graph communities rather than isolated chunks. ➔ The Trust Gap: How explicit entity-relationship mapping provides a "hard" logic layer to ground the LLM’s "soft" probabilistic reasoning. Intelligence is the engine & Structure is the guardrail. #GraphRAG #LLMSystems #TrustworthyAI

171

7 Comments

Search is evolving from keyword matching to intent understanding. Modern systems combine semantic embeddings with traditional search for higher precision. Seeing how AI-enhanced search improves user experience at smartData Enterprises Inc. has been insightful. Amit Ranjan Srivastava #AISearch #SemanticSearch

2

An interesting perspective on India’s role in the global AI ecosystem. This Times Special article highlights how India has become one of the most important markets for AI adoption — not just in terms of users, but in how AI products are validated and refined at scale. From a go-to-market standpoint, a few points stand out: • India offers rapid feedback across pricing, onboarding, and real-world usage • Distribution partnerships are lowering adoption barriers significantly • Products that succeed here tend to be more resilient and adaptable globally India is increasingly less of a late-stage market and more of an early signal for what works in AI at scale. For teams building and taking AI products to market, this is a reminder that GTM strategy and product strategy are now deeply intertwined, especially in markets like India. Sharing the article below — worth a read. #GoToMarket #AI #India #ProductManagement #SaaS #AIAdoption

15

1 Comment

How LLMs Work Is Not Magic Anymore A few weeks back I had taken a deep dive into the world of Large Language Models (LLMs) — and it’s been truly fascinating to unpack what goes on under the hood of models like GPT, Claude, and others. What once felt like “magic” now feels like brilliant engineering and elegant math. Through studying the Transformer architecture, I’ve developed a clear understanding of how these models actually think and learn to generate coherent, human-like text. Here’s what I’ve explored in detail 👇 🔹 Encoder–Decoder Architecture (Seq2Seq Models) – Understanding how data flows through the encoder to create context-rich representations, and how the decoder turns that into meaningful output. 🔹 Tokenization & Embeddings – How text is broken down into tokens and represented as dense vectors in a high-dimensional space. 🔹 Self-Attention Mechanism – The core innovation that allows models to understand relationships between words across long contexts. 🔹 Transformers – The building blocks of modern LLMs, replacing recurrence with attention for parallelism and scalability. 🔹 ReAct Framework – Combining reasoning and action to enhance the model’s problem-solving and decision-making capabilities. 🔹 Mixture of Experts (MoE) – The concept of dynamically routing tasks to specialized “expert” subnetworks to improve efficiency and performance. The deeper I go, the more I appreciate how systematic, scalable, and elegant these architectures are. ✨ What’s most exciting is that LLMs are no longer just black boxes to me — I can now understand and explain the “why” behind their behavior. #AI #MachineLearning #LLMs #Transformers #DeepLearning #NLP #ArtificialIntelligence #LearningJourney

2

AI isn’t about replacing people — it’s about unlocking productivity and enabling humans to focus on higher-value work. Hear more on this space from Arundhati Bhattacharya.

2

“Origin of Quantum Decoherence In theory”: A review by Rajeev Singh Sisodiya Quantum decoherence originates from the unavoidable interaction between a quantum system and its surrounding environment, which disrupts the coherent evolution described by the Schrödinger equation. This process explains the transition from quantum to classical behavior and represents the primary obstacle to practical quantum computation. The fundamental mechanism is rooted in quantum entanglement. In theory, the state of an isolated system evolves unitarily as $iħ ∂ψ/∂t = Ĥψ$. However, perfect isolation is impossible. Any real system interacts with degrees of freedom in its environment—photons, phonons, stray fields, or measuring apparatuses. This interaction entangles the system's state with the environment's states. As a result, phase information between the system's quantum superposition states is not lost but rather delocalized and dispersed into the vast number of environmental degrees of freedom. From the perspective of the system alone, this manifests as a loss of quantum coherence, effectively converting quantum superpositions into classical statistical mixtures. This process, decoherence, typically occurs on timescales orders of magnitude faster than time T₂ often << energy relaxation time T₁. The critical importance of decoherence lies in its role in bridging the quantum-classical divide and its direct impact on quantum technologies. It provides a physics-based explanation for the absence of macroscopic quantum superpositions, as large objects decohere almost instantaneously due to their strong coupling to the environment. In quantum computing, decoherence is the principal source of error and noise. The coherence time—the duration for which a qubit maintains its quantum state—directly limits the complexity of algorithms that can be executed reliably. For instance, in superconducting qubit architectures, decoherence times constrain gate fidelity and circuit depth. A superconducting [transmon qubit] may have a dephasing time (T₂) on the order of tens of microseconds, setting a hard upper bound on the number of sequential operations possible before information is corrupted. Overcoming decoherence is a multi-front endeavor, as no single solution exists. Current strategies focus on mitigation, correction, and intrinsic resilience:

4

https://lnkd.in/gwr_aAYk 🚀 Unlocking the Power of LLMs: It's All About the Ecosystem! Think of a Large Language Model (LLM) like a database: you send a query, you get a response. But unlike deterministic, structured databases, LLMs are stochastic, unstructured, and can sometimes hallucinate. The real magic? The ecosystem around the model that makes it reliable and impactful. Here’s a simple AI adoption strategy: treat the LLM as a black box that can produce brilliant responses or hallucinate. Control outcomes by: Pre-processing queries: Craft clear intent with cumulative context and guardrails (content moderation, jailbreak detection, sensitive data filtering). Post-processing responses: Evaluate and re-rank outputs for relevance, safety, and accuracy. Building a feedback loop: Refine prompts and learn from outputs to improve accuracy over time. The full AI interaction loop: User → Query → LLM → Response → Feedback → Next Query Each step is an opportunity to enhance safety, context, and quality. From human or AI agents wielding tools (memory, retrieval, orchestration) to red teaming and mechanistic interpretability, a robust LLM pipeline ensures reliable, responsible AI. How are you building your AI ecosystem? #AI #LLM #MachineLearning #PromptEngineering #ResponsibleAI #GenerativeAI #NLP #ContextEngineering

10

1 Comment

🇮🇳 Sarvam open-sourced AI models built for India. This could be a defining moment for India’s AI ecosystem. Sarvam AI enough to run high-throughput applications at scale. Think multilingual assistants, enterprise copilots, and customer service systems. Sarvam-105B • Mixture-of-Experts architecture • 128K context window • built for deeper reasoning, coding, and agentic workflows Despite having 105B parameters, it activates only ~9B parameters per token, enabling large-model capability with manageable compute. A technically interesting detail: their RL pipeline ↳ Beyond pretraining and supervised fine-tuning, Sarvam invested heavily in reinforcement learning alignment. ↳ Their RL system uses an asynchronous GRPO architecture, separating generation, reward computation, and policy updates to keep GPUs fully utilized during training. ↳ Instead of relying on traditional KL-regularized RL, they use a group-relative objective and reward shaping to encourage: • structured reasoning • concise answers • correct tool usage This helps stabilize RL training for large Mixture-of-Experts models. Why this matters ↳ For decades, India built world-class software on global platforms. ↳ Now we’re starting to see AI infrastructure built in India, for India. ↳ Infrastructure that could power: • multilingual citizen services • voice-first AI for the next billion users • enterprise copilots across regional languages • AI systems that work beyond English Sarvam’s open release of 30B and 105B models could be an early signal of that shift. A technical milestone — and possibly an important step in India’s AI journey. https://lnkd.in/dbpQWFkT #IndiaAI #OpenSourceAI

15

1 Comment

The next decade of AI won’t be decided by who builds the biggest model. It will be decided by who deploys AI reliably at scale. 🇮🇳 India is positioning itself to lead. This isn’t another AI announcement. It’s a signal. India is moving from AI curiosity to AI capability at scale. What we’re witnessing now is a shift where generative AI is no longer confined to labs, pilots, or slide decks — it’s being engineered directly into enterprise operations. Why India Is Uniquely Positioned to Win in AI • One of the world’s deepest engineering talent pools • Enterprises ready to leapfrog legacy tech with cloud-native AI • A cost-efficient market that demands automation at scale • Strong alignment between cloud, data, and AI adoption • Growing focus on production-ready, secure, enterprise GenAI India doesn’t just adopt technology. It industrializes it. The Real Opportunity Ahead This is not about chatbots. It’s about: • AI running IT operations • AI modernizing telecom and networks • AI accelerating enterprise decision-making • AI embedded into the “business stack” India is quietly positioning itself to be that place. The AI future isn’t knocking. India is opening the door. #IndiaAI #GenerativeAI #EnterpriseAI #AIAtScale #DigitalIndia #FutureOfWork #TechTransformation

3

In recent days, there has been quite a bit of debate around context graphs spurred by a thought-provoking article by Jaya Gupta and Ashu Garg! In our blog (link below in comments), Hui Zhang and I take a "first-principles" look at what the conversation got right and what's missing from a 𝙢𝙖𝙧𝙠𝙚𝙩, 𝙩𝙚𝙘𝙝𝙣𝙤𝙡𝙤𝙜𝙮, 𝙖𝙣𝙙 𝙡𝙤𝙣𝙜 𝙩𝙚𝙧𝙢 𝙚𝙫𝙤𝙡𝙪𝙩𝙞𝙤𝙣𝙖𝙧𝙮 perspective. To truly unlock the potential of future agentic systems, we identify three challenges and opportunities along these lines: 1. Explicitly tackle the needs of *𝗰𝗼𝗻𝘀𝘂𝗺𝗲𝗿-𝗰𝗲𝗻𝘁𝗿𝗶𝗰 𝗮𝗴𝗲𝗻𝘁𝘀* and not just enterprise business automation  2. Create technical foundations for *𝘀𝘁𝗮𝘁𝗲𝗳𝘂𝗹 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺𝘀* at scale to model behavioral patterns and trajectories of users, agents, and the interactions. 3. Move beyond trying to mimic human workflows and try to achieve *𝘀𝘂𝗽𝗲𝗿𝗵𝘂𝗺𝗮𝗻 𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻* or hit the "AlphaGo" moment (i.e., the AI system can generate an action to achieve an optimized outcome unconstrained by human compute/memory/expertise!) Read the blog linked in the comment below and let us know what you think!

91

9 Comments

We are partnering with early-stage venture firm Activate to identify and empower promising AI startups in India, reinforcing a locally driven innovation pipeline aligned to our computing ecosystem. Activate joins top funds like PeakXV & Accel in the NVIDIA Inception program. Learn more.

10

1 Comment

Something has quietly shifted in how we think about AI systems over the past year. Training used to dominate the conversation. Bigger models, more GPUs, better data. But the frontier has moved. The hard problems now live at inference time, and the engineering community is only beginning to catch up with that reality. A few things have made this obvious recently. Reasoning models like DeepSeek R1 and the o-series fundamentally changed the math. Longer chains of thought, more tokens per request, unpredictable output lengths. The serving infrastructure we built for short completions just does not hold up under those conditions. Throughput assumptions fall apart. KV cache pressure spikes. Latency variance becomes brutal. Disaggregated prefill and decode is another one. Separating the prefill phase from the decode phase onto separate hardware pools is no longer just an optimization trick. It is becoming a serious architectural pattern for anyone running inference at scale. Chunked prefill, speculative decoding, and prefix caching are no longer niche topics in research papers. They are production decisions that affect cost, latency, and user experience every day. And then there is the cost conversation. With reasoning tokens multiplying and GPU costs not dropping as fast as people hoped, token economics is becoming a real discipline. Semantic routing sits right at the intersection of that, matching requests to the right model based on actual intent rather than routing everything to the biggest, most expensive one. This is the layer of the stack that matters most right now. Not the model weights. The serving decisions wrapped around them. On 14 March, the vLLM community is gathering in Pune for an afternoon focused entirely on this. Red Hat, NeevCloud®, and Hewlett Packard Enterprise are hosting talks and a hands-on workshop covering vLLM internals, semantic routing, and what production-grade inference actually looks like when traffic is real and budgets are finite. If you are in Pune and working anywhere near inference infrastructure, model serving, or LLM systems in production, this is worth your Saturday afternoon. Spots are limited, and registration closes 24 hours before the event. RSVP here: https://lnkd.in/gBStRSEE #vLLM #Inference #LLM; Prasad - Rohit - Amit - Neha - Shamsher

26