Tejas Sarma

If you are someone pushing boundaries of LLM interaction, improving prompt engineering strategies, and most importantly building more robust AI application, this paper on agentic context engineering could be a good read for you !! https://lnkd.in/gG97xZGk

39

1 Comment

A Thought Exercise : Context Engineering with LLMs is analogous to Feature Engineering in Classical ML What are the similatieties and differences ? Share in the thread #AImusings courtesy: Jim Hong :)

28

9 Comments

🏥 How We Cut Pipeline Build Time by 65% Using GenAI + AWS in Healthcare Data Engineering In our healthcare data ecosystem, we handle terabytes of clinical and claims data daily — everything from Epic HL7 feeds to pharmacy transactions and patient encounter data. Initially, our data engineering team spent days building and testing ETL workflows. Every schema change or data mapping meant hours of PySpark and SQL rework. The process was accurate, but not scalable — especially as our cloud infrastructure on AWS (S3, Glue, Lambda, and Redshift) kept growing. That’s when we introduced GenAI-assisted data engineering, tightly integrated into our AWS stack. Here’s what that looked like 👇 🔹 Before: Manual ETL creation using AWS Glue and PySpark scripts. Repetitive SQL transformations and schema mapping across HL7/FHIR models. Code reviews + test validation cycles took ~3 days per pipeline. Frequent performance tuning for Redshift queries and data partitioning. 🔹 After (with GenAI + AWS integration): Used GenAI copilots to auto-generate PySpark transformations and Glue jobs. AI suggested S3 bucket partitioning strategies and optimized join logic for Redshift. Generated validation scripts for schema drift detection in Athena. Automated pipeline documentation pushed to AWS Glue Data Catalog. 📊 Impact: Pipeline build time reduced from ~30 hours to 10 hours. Query performance improved by ~40% due to AI-suggested optimizations. Data validation accuracy up 45%. Engineers now focus on business logic and compliance workflows instead of low-level data plumbing. 💡 Why this matters in healthcare: Data accuracy, traceability, and compliance are everything. By using GenAI inside AWS-native tools, we’re accelerating development without compromising HIPAA or governance standards. This combination of AI + Cloud-native engineering is helping us modernize faster — moving from reactive data management to proactive insight generation. 💭 Curious to know — Are you or your teams experimenting with GenAI for data engineering in regulated industries like healthcare or finance? How are you balancing automation, cloud scalability, and compliance? #Healthcare #DataEngineering #AWS #GenAI #ETL #DataPipelines #AIinHealthcare #BigData #AWSGlue #Redshift #S3 #DataOps #HL7 #FHIR #OpenToWork #JobSearch #CareerGrowth #HiringNow #DataEngineerJobs #TechCareers #JobSeeker #GenAI #AIinDataEngineering #MachineLearning #Automation #AIAssistedEngineering #TechInnovation #DataDriven #DigitalTransformation #CloudData #FutureOfData #TechLeadership #C2C #C2H

12

Personal Experience on Using BharatGen's PARAM 1 – A Step Toward Truly Inclusive AI Over the past few days, I've had the opportunity to explore PARAM 1, BharatGen's Indic-scale bilingual foundation model and I have to say, it's one of the most refreshing AI experiences I've had in a while. What stood out to me? -> True bilingual understanding: I could interact naturally in both English and Hindi sometimes even blending the two. Unlike many global LLMs, Param 1 didn't flinch. It understood the context, the tone and the nuance. -> India-centric training data: You can feel the difference. The responses feel grounded, culturally relevant, and more aligned with everyday Indian realities. -> Open and accessible:As a developer and AI enthusiast, having free access to a model trained on 5 trillion+ tokens with 2.9 billion parameters - that's a big deal. Whether it’s education, health, governance, or content creation the applications are massive. For us Indians, this isn't just a model - it's a foundation for inclusive digital transformation in India. If you're building in the Indic AI space, I highly recommend checking it out. Explore the model here: https://lnkd.in/dJVcTBCF #AIforBharat #BharatGen #Param1 #MultilingualAI #DigitalIndia #OpenSourceAI #InclusiveTech #LLMs #HindiLLM #AIExperience #IndianAI

12

🚀 Demystifying How Large Language Models (LLMs) Actually Work We often hear about pre-training, fine-tuning, and reinforcement learning when it comes to LLMs — but what do these stages really mean? Let’s break it down in simple terms: 🔹 Pre-training Think of this as the foundation. We have two files: A run file (small, with key instructions). A parameter file (huge, compressed, requiring GPUs to process). This stage teaches the LLM the structure of language by exposing it to massive datasets. 🔹 Fine-tuning Now we feed the LLM questions and answers. This allows it to specialize and learn from more focused examples. 🔹 Reinforcement Learning Here’s where feedback loops in: We ask the LLM questions. It generates answers. We rate those answers, and over time, the LLM gets better at giving responses that align with human expectations. 🧠 Why Transformers Matter At the core of all this is the transformer architecture — a neural network that makes calculations based on weights (numbers). Since text is messy, we break it into tokens (word pieces). Every prediction is basically the model calculating: “What’s the most likely next word?” But there’s a catch — once the token limit is reached, the LLM forgets older context. That’s why asking the right questions becomes so important which is called prompt engineering. ✨ Key Takeaway LLMs aren’t magic — they’re math, architecture, and feedback loops working together. The quality of questions we ask directly shapes the quality of answers we get. 👉 What do you think: is better data or better questions more important for improving LLMs? Pls share, like or comment to make sure the post reaches as many people as possible :-).

13

2 Comments

𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗶𝘁 𝗺𝗲𝗮𝗻 𝘁𝗼 𝘀𝗰𝗮𝗹𝗲 𝗰��𝗻𝘀𝘂𝗺𝗲𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝘀 𝗶𝗻 𝗮𝗻 𝗔𝗜-𝗳𝗶𝗿𝘀𝘁 𝘄𝗼𝗿𝗹𝗱? Dived deep into this at the roundtable at Bits n Atoms, hosted by The Product Folks last night — and the conversations stuck with me. Scaling consumer products in an AI world demands thinking across multiple dimensions, and these are just some of the lenses through which our group explored it.  • 𝗣𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘀𝗮𝘁𝗶𝗼𝗻 — not just smart recommendations, but designing for one and actually impacting behaviour change  • 𝗧𝗿𝘂𝘀𝘁 — especially for this new generation of AI-native products, where UX becomes critical to adoption  • 𝗖𝗼𝘀𝘁-𝗲𝗳𝗳𝗲𝗰𝘁𝗶𝘃𝗲𝗻𝗲𝘀𝘀 — using AI to boost efficiency and also designing AI systems that can scale without exploding infra or model costs  • 𝗠𝗼𝗻𝗲𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻 — whether through subscriptions or new models that make sense for AI-first experiences  • 𝗗𝗶𝘀𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻 & 𝗚𝗧𝗠 — this differentiation matters more than ever, in a world where building products is getting faster and cheaper everyday  • 𝗔𝗰𝗰𝗲𝘀𝘀𝗶𝗯𝗶𝗹𝗶𝘁𝘆 — across regions, languages, devices, and modalities. If it only works for the few, does it really scale? We’re in an interesting moment — where product thinking is expanding, not just because of what AI can do, but because of the questions it’s forcing us to ask. Loved the sharp, honest, and wide-ranging perspectives from everyone Shreyanshi C. Kushal Khandelwal Ashish Jain Madhav Bhartia Tanmay Saxena Pranav Agrawal Sachin Kamkar Shashank Raghavendra Nikhil Ankan Huge shout out to Abhay Jani Aditya Mohanty Suhas Motwani for driving these thoughtfully curated and highly impactful events!

92

6 Comments

Would models know more about Indian food in Hindi and Turkey’s history in Turkish? Does the language of a question affect an LLM’s answer? ✨Yes! ✨ Ishika Agarwal and I are excited to announce our newest preprint in which we explore “Language Specific Knowledge (LSK)”. By carefully selecting the language of the question asked to the language model, we can improve model reasoning and accuracy by 10%. We also reveal some interesting relations between models and languages. Our method is called LSKExtractor. First, we map out the LSK-to-language map. We notice that certain topics are best answered in certain languages. Next, during testing, we use the LSK map to select which language to translate questions into. We find the semantically closest query to the test question in our map and translate the test question to that language for the final output. We test this out on a bunch of models and various datasets, and find an average 10% increase in model performance. We also show which languages tend to map to the most topics: Chinese, German, French and Arabic seem to be common. Interestingly, gemma-3-1b has the most knowledge in Vietnamese! Would love to hear the community’s thoughts on this :) We are grateful to our advisor Dilek Hakkani-Tur, special shoutout to Sagnik Mukherjee, Priyanka Kargupta, and other ConvAI@UIUC lab members 🎉 📄Paper: https://lnkd.in/gzyR8A59 💻Code: https://lnkd.in/gJs_ixuf

258

9 Comments

Wrote an article on Byte Pair Encoding (subword Tokenization) and how it works which was potentially used in the the earlier versions of GPT / traditional Language Learning model paradigms. Took a sample of Hindi Language corpus text data to showcase the same. Article below - https://lnkd.in/gCu7Bgpf

33

Why India Needs Its Own Diagnostic AI Model Most diagnostic AI today is imported — trained on Western data, languages, and clinical contexts. But healthcare in India looks, sounds, and behaves differently. Different epidemiology. Different clinical workflows. Different constraints. That’s why we built NEXUS — India’s first Diagnostic Reasoning Model. 👉 Neural Evidence Extraction & Uncertainty Synthesis — designed to reason clinically, in India’s languages, across India’s realities. Proud to build this in partnership with IIT Hyderabad | CfHE | Google Cloud for Startups. Because India doesn’t just need AI in healthcare. It needs Healthcare AI that speaks India. #AI #HealthcareAI #Diagnostics #MadeInIndia #Innovation #Aivana #NEXUS #ClinicalAI

27

🔍 Function Calls vs Model Context Protocol (MCP) in Agentic AI 🤖 Function calling brought structure to LLM interactions — enabling models to trigger tools, retrieve data, and execute tasks. But in agentic systems, where LLMs act as thinking agents rather than reactive bots, function calls alone don’t cut it. Enter: Model Context Protocol (MCP) 🧠 🔹 Function Calls = “Here’s a tool. Use it.” 🔹 MCP = “Here’s context, memory, history, toolset, persona, and environment — go accomplish a goal.” Why MCP matters: • ✅ Context-rich reasoning — gives the model structured access to user goals, memory, history, and tool capabilities. • 🔄 Multi-turn continuity — allows agents to persist state and adapt plans over long tasks. • 🛠️ Tool orchestration — lets the model choose and combine tools intelligently rather than just calling one-off APIs. • 📦 Modularity & Portability — MCP creates a standardized, modular interface for managing agent inputs and outputs — beyond raw JSON and ad hoc hacks. Think of function calls like giving an intern a list of API docs. Think of MCP like onboarding a smart assistant with access to calendars, tasks, Slack, past conversations, and team goals — and letting them figure out what to do next. Agentic AI needs structure beyond prompts. MCP is that structure — enabling autonomy, reasoning, and reusable frameworks at scale. We’re moving from “chat with AI” to “collaborate with AI agents.” #AgenticAI #MCP #LLMs #AIInfrastructure #ToolUse #AIContext #AIArchitecture #GenerativeAI #FunctionCalling

38

2 Comments

When math draws a hard line for large language models A recent research paper by Vishal Sikka and his son Varin Sikka takes a very different approach to evaluating large language models. Instead of benchmarks and demos, it uses mathematics. The authors show that LLMs operate within a fixed computational limit. Once a task requires more computation than the model can perform during inference, two things happen: the model cannot reliably solve the task and cannot verify correctness either. Incorrect output becomes unavoidable. This applies directly to many “agentic AI” scenarios - long planning chains, multi-step decision making, global optimization, and autonomous workflows. The paper does not argue that LLMs are useless. Quite the opposite. Within their computational domain, they are incredibly powerful tools. What it shows is that some classes of problems sit permanently beyond what transformer-based models can handle, no matter how much data or training we add. AI can become a phenomenal accelerator - for code, analysis, automation, and knowledge work. At the same time, treating it as a universal reasoning engine or a fully autonomous problem solver creates real risk. The research is a good reminder that understanding the limits of a tool is just as important as admiring its strengths. You can read about in more detail here: https://lnkd.in/dqQBQZxd

12