I'm writing this from Himshikhar, a two-month residential program at IIT Mandi where my track is Agentic AI. The days are dense: professor lectures, labs, industry-mentor sessions, daily quizzes, and a mini-project. Mine is on structured LLM workflows, and it's quietly reshaped how I think about building with language models.
The setup
We covered the fundamentals fast — deep learning basics, then the architecture zoo (FCNN, VAE, RNN, LSTM, transformers) — before getting to the part I came for: LLM reasoning, retrieval-augmented generation, and agentic frameworks. The daily quizzes are annoying and extremely effective; they don't let fundamentals stay fuzzy.
From answers to steps
The mental shift the program keeps pushing: an LLM call is not an answer, it's a step in a workflow you design. A structured workflow wraps the model in tools, retrieval, checks, and explicit control flow, so the system is composable and inspectable instead of one opaque prompt you pray over.
A bigger model makes the single call better. Structure makes the whole system dependable. Those are not the same lever.
Reliability is engineering
The recurring theme: agentic reliability is an engineering problem, not just a prompting one. You get there with retrieval that grounds the model, steps that can be checked, and fallbacks when a step fails — not with one more clever instruction. RAG stops being a buzzword once you've watched a grounded answer beat an ungrounded confident one.
What I'm carrying back
I want to take structured-workflow patterns back into ExplainMyRepo, where the hard part was always retrieval and framing rather than generation. And I'll probably write a proper retrospective once the program ends — this is the in-progress version.