Cohort 03 · Live · Weekends only
5 September – 27 September 2026 · Weekend morning, 11:00 IST
Four weekends, weekends only. Ten hours of instructor-guided building, six hours of live coaching on the agent you build yourself. You leave with something running a real task from your own work — not a certificate for watching videos.
You've wished for a second version of yourself. This is the closest thing that actually ships.
Flowise + n8n · Free LLM API keys · No prior AI experience
A starter kit you can build as your own
Same skeleton every build follows — arrives → agent reads → grounded → gate → output.
The gap
Chatting with a model saves you ten minutes. An agent that runs unattended every morning saves you the task. The difference is not intelligence — it is plumbing, and plumbing is now drag-and-drop.
Brainstorming with a model is real leverage, until you close the laptop. There's no trigger, no backlog, no memory, no record, and nothing moves while you're in a meeting or on a flight. You learn to turn a prompt into a process that runs without you in the room.
Most agent content is written for people who read Python. This course is built the other way round: you'll understand embeddings, tool calls and context windows properly, and you'll never open a terminal to do it.
An agent that works in the room and breaks on Monday is worse than nothing. Error branches, approval gates, evaluation sheets and cost ceilings are taught as part of building — not as an afterthought.
Who joins
Pick the one closest to you.
“The thinking is the only part that needs me. Everything after it is just execution I don't have time for.”
Founders and business leaders join because the agent becomes a second version of them for the repeatable work: the follow-ups, the filings, the status updates, the chase. You design the decisions; the agent executes them.
You walk out with
How it runs
Closer to a group class than a lecture. Everyone follows the same build on screen, then repeats it on their own use case. In the coaching sessions a handful of those attempts get worked through live — and the rest of the room learns from the fix.
Five sessions, all following one agent as it gets built end to end. You are not watching slides — your canvas is open and you build the same thing at the same time. Each session adds a layer to the same agent, so by the last one it is a real, deployed system.
Every session ends with a handout: the exact steps to repeat on your own use case.
You apply those same steps to a use case of your choosing, or one of six starter kits. Where you get stuck, you send a short note ahead of the session. The instructor triages them and works the common failures live, on real screens.
One flow gets fixed on screen. Thirty people learn the fix.
Learning session
Build the reference agent alongside the instructor, layer by layer.
Your build
Repeat the same steps on your own use case, in your own time.
Coaching session
Send what broke. Watch it get fixed live, along with everyone else's.
Repeat
Four weekends, until your agent runs on its own.
The curriculum
Rather than five shallow demos, the class builds a single agent — the Desk — and each session adds one layer to it. Documents and requests arrive, it reads them, answers from your own reference material, files what it can and escalates what it can't. Every layer transfers directly to whatever you build on your own.
5 instructor-guided sessions · 2 hours each · weekend morning, 11:00 IST
What an agent actually is, how a language model behaves, and why most people's prompts fall apart the moment something depends on them. The longest hands-on stretch of the course, because everything after this assumes it.
Covered in session
The Desk gains · live
Layer one. A request arrives as messy text; the Desk returns owner, category, deadline and urgency as clean structured fields — the same way every time.
Your handout
Pick your use case and your starter kit. Build the same extraction layer on your own inputs. Note every case where the output came back wrong.
Your own build
The class builds the Desk together. Between sessions you build your own version of it — same five layers, your domain. Start from one of these six kits, each shipped with sample data and a template so nothing stalls for want of a spreadsheet. Or bring a process from your own job; we'll scope it together early on so it's finishable inside the cohort.
Same skeleton every time — something arrives → agent reads → grounded → gate → output
Invoices arrive as PDFs in a mailbox and a Drive folder. Your agent reads them, pulls GSTIN, HSN, taxable value and tax heads, matches each against the purchase register, and posts a daily exception report of everything that doesn't tie out. Nobody opens a PDF manually again.
Invoice lands in mail or Drive
Extract GSTIN, HSN, taxable value, tax
Match against purchase register
Flag mismatches over tolerance
Daily exception report to finance
Screen-record your agent doing its job and talk over it. No live presentation, no scheduling, no stage fright — and you keep the video.
Does it run unattended? Is it grounded in a real source? Does it fail safely? Do you know what it costs per run?
A rubric score and notes on every submission, not just the strongest. Certificate on a pass. Best builds published to the cohort gallery.
The tools
We teach the pair together because they solve different halves of the problem — and because connecting them is what turns a clever demo into an operating system for your work.
A visual builder for LLM applications. You drag nodes onto a canvas to compose chains, retrieval pipelines and multi-agent flows, then publish them as a chat widget or an API endpoint.
A workflow automation platform with several hundred integrations. It gives your agent triggers, credentials, access to the tools your business already runs on, and a full log of every execution.
Sayantan Dawn
Data Science & Generative AI Educator · Risk Modelling Specialist
Who's teaching
Ten years across analytics, credit risk modelling and generative AI at Citi, JP Morgan, Tiger Analytics and BRIDGEi2i. Founder of Last Bench, where he has mentored more than 500 students and professionals into roles at Citi, HSBC, Barclays, JP Morgan, Genpact and EXL.
Every session is delivered live, with a cohort small enough that you can interrupt. Nothing is pre-recorded, nothing is read off a slide, and every build happens on screen in real time — including the parts that break.
Years in analytics & GenAI
Mentored into roles
Institutes taught at
Workshops at IIT Kanpur · IIT Bombay · ISI Delhi · IIFT · BHU · Calcutta University
Before you start
Shorter than most lists, deliberately. There is no Python prerequisite, because there is no Python.
Not needed: coding, a CS degree, a paid API account, or prior AI experience.
On the other side
Anyone can attend. The certificate is issued only once the capstone clears the rubric — so it means something specific.
Enrolment
No upsells, no premium tier, no separate charge for the recordings.
Cohort 03 · Live seat
₹3,999
+ 18% GST · per person
Cohort 03 runs 5 September – 27 September 2026 · begins Sat 5 September, 11:00 IST
Thirty seats, because six hours of coaching split thirty ways is the honest limit.
Sixteen hours live, which puts an hour of it at about the price of a pizza. GST invoice issued against your company GSTIN, so input credit can be claimed.
What the fee covers
Nothing else to buy
Flowise and n8n are open source, and their free tiers cover everything we do. Free LLM API keys are issued to you in week one. Your spend beyond the course fee is zero.
FAQ
No — that's the point of the gate and the evaluation layers. You'll build one agent you actually trust to run unattended, and you'll know what it costs per run. Most founders in the room build a delegation agent for themselves first, then hand the same pattern to their ops or EA to extend.
No. Every agent in this course is built by dragging nodes onto a canvas and connecting them. You'll learn to read a data structure and write instructions in plain English — but you will not write a single line of code, and no session assumes you've ever opened a terminal.
Nothing is ever scheduled midweek. Eight live two-hour sessions across four weekends — five instructor-guided learning sessions (10 hours) and three coaching sessions (6 hours), all on weekend morning, 11:00 IST. Beyond that, budget roughly two hours a week on your own build, whenever suits you. The full session calendar goes out with your joining pack.
Come anyway — that's where a lot of the learning happens. Each one opens with the same pattern rebuilt in a different domain, then Sayantan works through the failures the cohort actually hit. Watching someone else's broken retrieval get fixed is often more useful than fixing your own, because you see the diagnosis rather than just the answer.
Yes, and professionals usually should. We'll scope it together early on so it's realistically finishable inside the cohort. The only difference is materials: the six kits ship with sample data and a template, and your own use case won't. Coaching access is identical either way.
Nothing. Flowise and n8n are open source and their free tiers cover everything we do, and free LLM API keys are issued to you in week one — no credit card, no prepaid credits, no bill arriving later. Beyond the course fee, your spend is zero.
Yes, and you have an advantage you may not expect. Agents are only useful when someone genuinely understands the process being automated — a CA student understands reconciliation better than most engineers do. Domain knowledge is the hard part. The canvas is the easy part.
You submit a three-minute screen recording of your own agent within a week of the final session — no live presentation, no scheduling. It's scored against four published criteria and the certificate is issued on a pass, with written feedback either way. The recording itself is the better credential: you can send it to a recruiter or a manager.
Yes. Add your company name and GSTIN on the enrolment form and the invoice is raised in your organisation's name before the cohort starts, so input credit can be claimed. For teams of five or more, mention it in the form and we'll set up a single consolidated invoice.
Everything is recorded and shared within 24 hours, and the coaching sessions exist precisely to catch you up — so a missed learning session is recoverable. That said, this is built around live building. If you expect to miss more than two of the eight, the next cohort is a better fit.
Any Windows, Mac or Linux laptop with 8 GB of RAM and a stable connection. If you'd rather not install anything, both Flowise and n8n can be run entirely in the cloud and we'll walk through that setup in the setup call before the cohort opens.
Attend the first learning session in full. If it isn't what you expected, write in before the second session begins and the fee is refunded, no questions asked.
