Campus Immersion at IIT Madras

A Two-Day Memoir on Agentic AI, Architecture, and Operational Maturity, and interaction with eminent personalities in AI/ML research and development

Event details

Start date and time: 27 July 2026 at 3:45 am
End date and time: 28 July 2026 at 11:00 am
Host: IITM - Pravartak Technology Innovation Hub
Place: Raman Hall, IITM-Pravartak Campus, Chennai
Reading time: 15 min read
Published: 28 July 2026 at 11:44 pm

Prologue: Returning to Chennai

There is a distinct, unmistakable energy that fills the air in Chennai. Every time I set foot in this city, I am struck by its unique blend of deep-rooted academic reverence and restless technological ambition. This visit was special. I was returning not merely as an enterprise technology professional, but as a practitioner seeking to anchor months of intense exploration into a rigorous, formal framework.

The destination was Raman Hall at IIT Madras Pravartak, for the two-day physical campus immersion program: the culmination of the Professional Certificate Program in Agentic AI and Applications, delivered jointly by IITM Pravartak and Emeritus.

Over two packed days, surrounded by faculty, researchers, and an extraordinary cohort of peers spanning manufacturing, finance, healthcare, and software, we stepped beyond the industry hype to examine what it truly takes to architect, deploy, and govern autonomous AI systems in complex, real-world environments.

What follows is my account of those two days, revisited once more after the fact. A few of my original notes ran ahead of what was actually taught, and I have corrected those places rather than let a tidier version of events stand.

A note before I begin: wherever this account reflects a misinterpretation, an oversimplification, or a gap, that limitation is mine, not the faculty's. Dr. Madhusudanan, Prof. Srinivasan, Prof. Sandeep, Dr. Satyavratan, and Prof. Vasan Srini taught this material with far more nuance than a single attendee's notes can capture in real time. Where my first draft flattened that nuance, I have tried to correct it, and I remain glad to stand corrected on anything I have still gotten wrong here.

Day 1: First Principles, Knowledge Graph Reasoning, and Industrial Realities

Date: Monday, 27 July 2026

Morning Session: Beyond the Hype, First Principles and Model Strategy

The immersion began at 10:00 AM in Raman Hall with an opening masterclass by Dr. Madhusudanan. Right from his introductory remarks, the tone was set. We were not there to admire commercial wrapper demos, but to deconstruct artificial intelligence down to first principles.

Dr. Madhusudanan challenged us to step back from hype cycles and examine core system prerequisites:

Problem Definition vs. Data Quality: the most brilliant architecture fails if applied to an ill-defined problem or fed with uncurated data. Data remains the single most critical asset.
Domain Experience is Mandatory: AI models cannot validate themselves in a vacuum. Whether in steel manufacturing, banking, insurance, or healthcare, domain expertise must guide model validation.
Model Size and Cost Spectrum Optimization: bigger is not always better. We discussed token economics, latency trade-offs, and the temptation toward overengineering.

"Beware of overengineering. Beware of hidden cost bloat. Focus on solving the core problem using the leanest effective architecture."

A major thesis of the morning was the strategic push toward Domain-Adapted, Local LLMs, with a roughly 30-billion-parameter class of model (illustrative figure only, the right size depends on the hardware budget and the task on hand). While cloud-hosted mega-models dominate headlines, real-world enterprise transformation, especially in privacy-sensitive and air-gapped industrial sectors, increasingly leans on specialised, laptop-runnable models fine-tuned using parameter-efficient methods like LoRA and QLoRA.

Dr. Madhusudanan emphasized Sovereign AI. Frontier models hosted out of the US or China carry inherent cultural, regulatory, and geographical biases. Building sovereign capabilities, leveraging initiatives like India's 38,000-GPU compute infrastructure under the IndiaAI Mission, allows enterprises to maintain data sovereignty, manage organizational context, and align AI outputs with local regulatory mandates.

The above table shows two ends of a spectrum, not two competing camps. In practice, most serious deployments end up as a portfolio across both columns: frontier APIs for open-ended reasoning where nothing sensitive leaves the building, local fine-tuned models for anything touching proprietary or regulated data. That framing lines up with work I have been doing separately on dynamic model routing, and it is closer to what the session actually argued than a simple either-or.

Deep Dive: Where Standard RAG Ends and Knowledge Graphs Begin

As the morning progressed, we got into one of the most practical technical debates in modern enterprise AI: when does Retrieval-Augmented Generation reach its limits, and what fills the gap?

Standard RAG excels at retrieving textual fragments based on semantic similarity. But semantic vector distance alone cannot perform multi-step relational reasoning. In complex domains such as legal arbitration, regulatory compliance, or multi-layered industrial automation, retrieval by itself is not enough. Inference and structured reasoning are also required.

When evaluating legal frameworks (like India's BNS laws, organizational policies, or arbitration precedents), an LLM must understand ontologies, mapping relationships between decision nodes: Subject to Predicate to Object. Layering Knowledge Graphs alongside open-weights domain models (such as Gemma's 26B mixture-of-experts variant, for instance) significantly boosts explainability while keeping compute footprints manageable, especially when the vector layer underneath is still doing most of the retrieval work.

Afternoon Session: Agentic AI Real-World Use Cases and System Design

Following a networking lunch, Prof. Srinivasan Thiagarajan took the floor at 2:00 PM for a session focused on production-grade agentic architectures.

Dr. Srinivasan Thiagarajan during a live lecture

During this session, I had the chance to share my own perspectives on solving "TimeWarp" and latency synchronicity issues in multi-layered industrial automation systems, and on how structured retrieval pipelines unlock knowledge hidden across legacy database tables and enterprise applications.

Srinivasan walked us through real-world Document Intelligence pipelines like the one above, and it turns out to be the clearest illustration of the next point.

The Golden Rule of AI Application

Late in the afternoon, we arrived at a principle that every technology executive and solution architect should keep close: all Large Language Models perform probabilistic pattern matching.

Use Agentic AI for probabilistic domain problems, where context, ambiguity, and multi-variable reasoning exist.
For deterministic problems governed by hard rules, explicit math, or fixed transaction logic, lean on deterministic code instead.

The OCR confidence score is a probabilistic judgment made by the model, and that judgment triggers deterministic branching logic: automated pass above 85%, stratified sampling in the middle band, exhaustive human review below 60%. That is the real shape of most production agentic systems. AI handles the ambiguous judgment call, deterministic code handles everything downstream of it. The rule is not "AI or code." It is closer to: AI decides where judgment is genuinely needed, code executes everything else.

Evening Reflection and Schedule Pivot

Our original schedule called for an outdoor Research Park (RP) visit at 3:45 PM. Logistical adjustments and ongoing classroom discussions meant the tour was pushed to Day 2. Far from a disappointment, this extra time let our cohort linger in Raman Hall, trading views on local LLM fine-tuning, QLoRA hardware constraints, and industrial edge computing. Day 1 ended with strong professional bonds forged across industries, before we parted ways in the rainy evening.

Day 2: MLOps Maturity, Autonomous Workflows, and the Heart of IIT Madras

Date: Tuesday, 28 July 2026

Morning Session: Autonomous Capabilities, Beyond ChatGPT

Day 2 kicked off at 10:00 AM with Dr. Satyavratan Govindarajan leading a session focused on moving from conversational chatbots to autonomous agentic workflows.

Dr. Satyavratan guided us through a recap of agentic capabilities: orchestrating dynamic tool selection, connecting agent nodes to enterprise knowledge bases, writing custom plug-ins for office applications, and utilizing Reinforcement Learning (RL) for specialized domain adaptation. The goal of modern enterprise AI is not to generate conversational text, but to construct autonomous task execution loops where agents perceive inputs, reason over tools, execute sub-tasks, and validate their own outcomes under human supervision.

Technical Deep Dive: The MLOps Reality and Production Gap

At 11:15 AM, Prof. Sandeep delivered what was arguably the most sobering, reality-checking lecture of the immersion: MLOps and LLMOps.

Prof. Sandeep opened with a statistic that has become something of an industry cliche: roughly 87 percent of machine learning models built in enterprises never reach production. The failure rarely stems from model math. It stems from operational immaturity. Process maturity matters far more than raw model quality.

Prof. Sandeep deconstructed the CALMS Framework (Culture, Automation, Lean, Measurement, Sharing), a lens borrowed from classic DevOps and adapted here for machine learning, and pointed out three blindspots traditional DevOps teams face when they first encounter ML:

Data isn't versioned out-of-the-box: tracking code changes in Git is trivial; tracking multi-gigabyte training dataset evolution requires tools like DVC (Data Version Control) and Directed Acyclic Graphs (DAGs).
Correctness isn't binary: traditional code either passes or fails unit tests. AI models exhibit probabilistic distribution shifts; correctness degrades gradually over time.
Nothing retrains automatically without explicit loops: MLOps extends the paradigm from CI/CD (Continuous Integration / Continuous Deployment) to CT (Continuous Training). MLOps is a closed loop, not a linear path.

Personal Reflection on the MLOps Challenge

Coming from a business, product, and digital transformation background rather than a pure MLOps engineering setup, I felt genuinely challenged during parts of this deep dive into container orchestration, model drift detection registries (MLflow, Evidently, WhyLabs), and hyperparameter tuning frameworks (Optuna, Katib).

That discomfort is precisely where the real learning happens. Sitting with these engineering realities sharpens my ability to bridge business strategy with technical architecture, and gives me more confidence leading multi-disciplinary teams through a full AI transformation cycle rather than just the parts I already knew well.

Afternoon Milestone: Certificate Ceremony

The afternoon marked a proud milestone. During the valedictory session at 2:00 PM, we were formally presented with our certificates for the Professional Certificate Program in Agentic AI and Applications by Prof. Vasan Srini, Chief Project Officer of IITM Pravartak, and Dr. Madhusudanan.

Receiving this honor from leaders who are actively shaping India's deep-tech landscape made the months of hard work feel deeply rewarding.

Grand Finale: Pravartak Hub, Research Park, and Campus Bus Tour

With the formal certificates in hand, late Tuesday afternoon brought the culmination of the immersion, combining our facility visits and campus exploration into a grand final tour.

First, we visited the IITM Pravartak Innovation Hub, funded by the Department of Science and Technology, Government of India. Seeing dozens of deep-tech startups being incubated under one roof provided a clear window into India's emerging technology ecosystem.

Two standout indigenous initiatives particularly caught my attention:

Shakti Processor: India's open-source RISC-V processor initiative, building sovereign silicon hardware tailored for edge and secure computing.
Shakti DB: an AI wrapper built over a hardened PostgreSQL database core, showing how relational database standards can be augmented natively for vector and AI operations.

We then embarked on the full IIT Madras Campus Tour by bus, which included exploring the IIT Madras Research Park.

There is a serene, timeless rhythm to IIT Madras. As our bus navigated the tree-lined avenues, spotted deer grazed peacefully nearby while students cycled past residential blocks named after India's sacred rivers: Ganga, Godavari, Mandakini, Cauvery, and Saraswathi. It was a fitting end to the immersion, a rare space where cutting-edge deep-tech innovation meets quiet academic tradition.

Summary of Key Takeaways from the Immersion

Domain-adapted local models and frontier public models are not rival choices but complementary tools. The right architecture routes between them based on data sensitivity, cost, and task, not a blanket preference for one over the other. A roughly 30B-parameter local model is a useful illustration of the on-premise end of that spectrum, not a fixed rule.
Knowledge Graphs (Subject-Predicate-Object) do not replace vector-based RAG. They combine with it to fill the multi-hop reasoning gap that standard semantic retrieval cannot close on its own.
Reserve AI's probabilistic judgment for the problems that genuinely need it, and hand deterministic rules to deterministic code, often within the very same pipeline.
Process maturity dictates production success. MLOps is a continuous training loop (CI/CD/CT) built to combat model drift, and it borrows its cultural backbone (CALMS) from classic DevOps.

Epilogue: Building the Future

I left IIT Madras with a full notebook, renewed clarity, and a fantastic network of professional colleagues.

Earning this certification is not an endpoint; it is a springboard. As autonomous AI frameworks continue to mature, the responsibility falls on technology leaders to ground these advances in practical engineering, robust governance, cost discipline, and measurable business impact.

My sincere thanks to Prof. Madhusudanan, Prof. Srinivasan, Prof. Sandeep, Dr. Satyavratan, Prof. Vasan Srini, and the entire teams at IITM Pravartak and Emeritus for crafting such an exceptional learning experience.

The work continues: building robust, future-ready, and sovereign AI solutions for complex industrial ecosystems.