A hands-on course where you ship one working document agent — and finish knowing exactly how accurate it is.
One track for everyone · Bring your own documents · Nothing costs extra
You write code that reads documents, checks facts, and takes action — from the very first lesson.
You build an agent from scratch before touching LangGraph, so the frameworks never confuse you again. It's all the same loop underneath.
Scans with no text layer, stamps, rotated pages, tables that break across pages. The documents that break every tutorial you've followed.
Answer from the document, cite the exact page, and correctly refuse to guess. This is the hard, valuable part — and our day job.
Not "it seems to work." A field-by-field accuracy report you can put in front of a boss or an interviewer.
Every agent is a loop: reason → act → observe → repeat. Each module adds one capability on top — concept, live build, lab, checkpoint.
Make your first model call, get reliable structured output, then build a tool-using agent in about sixty lines of Python with no framework at all. Rebuild the same thing in LangGraph so you see exactly what the framework hides.
When the PDF is really a photograph. OCR, layout, reading order, tables, stamps and skew. Then structured extraction: define a schema, extract into it, validate the result, attach a confidence score, and build the path where the agent says "not in this document" instead of inventing an answer.
The module that earns the course its name. Your agent reads a document, looks something up in a second source, compares the two, writes a row to a sheet and drafts an email — with a human approval gate on anything it isn't sure about.
Build an evaluation set and measure accuracy field by field. Do error analysis on the failures that matter. Add guardrails, prompt-injection defence and on-prem privacy. Then everyone demos what they built, with their numbers on the slide.
Worth saying plainly, so nobody pays ₹3,999 for the wrong thing.
No beginner lane, no advanced lane. You just point your agent at your own documents.
See what the frameworks hide, and what breaks on a real scan.
Finish with a system and an accuracy number you can walk an interviewer through.
Sitting on a pile of PDFs. Every lab ships with a working template.
A setup notebook goes out a week before, so session one isn't lost to installs.
Join the next open cohort, or bring the same four modules in-house for your team.
IIT engineer, Master's in the US, ex-Amazon data scientist. At Grokking he builds the document-extraction pipelines behind agents in use at NHPC, Indian Oil and Maruti Suzuki. This course teaches that exact craft.
Basic Python — functions, loops, dictionaries — is enough, and a short primer is included if you're rusty. Every lab also ships with a working template, so you just need to be comfortable running a notebook.
Patterns first, framework second. You build by hand before touching LangGraph. Tools and data are exposed over MCP. We use a hosted model plus one open model you can run locally, so you can see the code is model-agnostic.
Model API usage across four weeks typically runs a few hundred rupees. If you'd rather not use a card at all, there's a local-model path for every lab.
Yes, and it's the better way to do the course. If they're confidential, use the local-model path so nothing leaves your machine. We also provide a public tender corpus if you'd rather not bring your own.
It's the hardest and most valuable part of enterprise AI: answering from real documents without hallucinating, every answer traceable to the exact page. It's also what Grokking builds commercially, so you learn it from production experience.
Tell us in advance and we'll help you catch up. Every lab ships with starter and solution notebooks and is designed to be completed asynchronously, and questions get answered in the group between sessions.
Yes — a two-day private version, on-site or private virtual, run on your own documents inside your own environment. Teams that go through it often move straight into a production build with us.
We help you show the work well — a write-up of what you built, a cleaned-up repository, and how to talk through it in an interview. We don't place people and we don't promise offers.
Starts 15 September 2026, runs four weeks, ₹3,999 with everything included. Tell us a little about you and we’ll send joining details.