Released an order with no human approver, 8 times out of 8.
The agent ran on the all-powerful service account IT had made "for the bot".
Acting as the signed-in user: 0 of 8. Session 7.
Build a production AI agent from scratch: the LLM, retrieval, tools over MCP, the agent harness, workflows and evals. Not a chatbot demo. An agent that does real work inside a 30-year-old business system, and gets past every trap it sets.
Formerly AI Agents for Enterprises. Same start date, instructor and fees, now twelve sessions.
With Dr. Rajat Dandekar · Saturdays and Sundays, 7:00–9:00 AM IST · 31 Oct to 6 Dec 2026
One agent, built piece by piece, until you understand every part of it.
You start with how a language model predicts the next token and finish with a deployed agent that acts as the right person, asks for approval when it must, and recovers when the system hangs. Each session adds one piece.
Why this course
We measured how, while building this course, on a sandbox ERP with real traps. Each failure below is one session's lesson, and the fix is what you build.
The agent ran on the all-powerful service account IT had made "for the bot".
Acting as the signed-in user: 0 of 8. Session 7.
A free tool loop quoted the certification policy, then ordered anyway when the user said yes.
A control graph with a certificate gate: 0 times, on a small fraction of the tokens. Session 5.
Redoing a job after a timeout, with no idempotency key, the agent ordered the same goods again and again in our drill.
Keyed tools with read-back: 50 of 50 ended with exactly one order. Session 7.
Asked for the company's stock, the bare model answered with a confident number every time. None of them came from the system.
Retrieval and tools give it the facts. Sessions 2 to 4.
And spent about 81,000 tokens on each one, tripping over codes and screen order.
15 tools, one per business activity: 29 of 30, on less than half the tokens. Session 4.
The ERP printed success for an order that would never reach the supplier.
Receipts read back from the system: 10 of 10 drafts caught, against 3 of 10 without. Session 7.
Runs on Gemini models through the course harness, October 2026; the logs ship with the course repository. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027.
The live app
Northbridge Cycles, an e-bike maker in Manchester, still runs on Ledgerline ERP from 1998. Its bike shops are complaining and a competitor confirms orders in two minutes. You are the engineer it hires.


Every company, person, code and number in the app is fictional.
Curriculum
Every idea is taught from first principles on the same running example, then added to the agent. Nothing is used before it is taught.
What you build
Each tool is a folder: a JSON schema the model reads, and the Python that runs it as the signed-in user, with a preview before every write and a receipt after.
Pick your capstone module
Instructor
Co-founder, Vizuara · PhD, Purdue University · B.Tech and M.Tech, IIT Madras
Rajat teaches reinforcement learning, AI agents and reasoning models. He built this course's sandbox, toolbox and harness, and ran every experiment whose numbers appear in it.
In the cohort you work through the design of an agent with him, implement its tools and controls, and test whether it does the job you intended.


Fees and enrollment · starts 31 October 2026
Or choose a bundle
Tell us your team size and the process you want an agent for. We reply by email, usually within a day.
Questions
Saturdays and Sundays, 7:00 to 9:00 AM IST (UTC+5:30), from 31 October to 6 December 2026. That is Friday and Saturday evening in the Americas: the first session is at 6:30 PM Pacific / 9:30 PM Eastern on Friday 30 October, and an hour earlier once US clocks change on 1 November. Across Asia it is Saturday and Sunday morning.
No. The agent you build takes real actions in a business system: it raises requests, waits for approvals, places and releases orders and records goods received, and it has to prove each step happened. The chat window is the smallest part of it.
You read and change real Python in every session, and you write your own tools, harness parts, workflows and evals in the capstone. You never start from an empty file: each notebook gives you working code and exercises with solutions.
The course runs on Gemini by default, and any OpenAI-compatible model (GPT, Claude and others) works through OpenRouter. Every notebook can replay our recorded model replies, so it runs on free Colab with no key. Live calls with your own key are optional.
The course was called AI Agents for Enterprises. What it teaches is how to engineer agentic systems, from the model up, so the name now says that. It starts on the same date with the same instructor and fees, and now runs twelve sessions: four more on the industry tools companies use. Anyone already enrolled keeps their seat.
Yes. Teams of 25 or more get 7% off. Use the form above with your team size and the process you have in mind, and we will send a quote and invoicing details. Private cohorts and work on your own systems are scoped separately.
In class everyone uses the course's fictional company and sandbox ERP. For your capstone you can model your own process, with synthetic or approved data.
It will be a working pilot you can inspect, test and extend: tools, harness, approvals, evals and a deployment. Taking it into a live system is a project of its own, and the course shows you what that project involves.
Twelve live sessions with Dr. Rajat Dandekar, starting 31 October 2026.