Stop watching AI videos. In one live session you'll understand how ChatGPT, Claude and AI agents really work, then write and run a working Python agent on your own laptop. Free accounts, no prior coding.
DateSun 25 Oct 2026Time10:00–11:30 AM ISTWhereLive on Zoom
The agent you'll build · real test run, 10 Oct 2026
Why this matters now
Everyone uses AI. Almost nobody knows what's inside it.
That gap is the difference between someone who pastes prompts and someone who builds with AI. This session closes it in one morning.
52%
of US workers are worried about AI's impact on their jobs.
Pew Research Center, 5,273 workers, October 2024
52%
of people who use AI at work are reluctant to admit using it for important tasks.
Microsoft & LinkedIn Work Trend Index, 31,000 people, 2024
90 min
is all it takes to go from "I use ChatGPT" to "I built an agent that does a task for me".
The plan for this session: six explain-then-do cycles
Who it's for
Built for where you are right now.
Same 90 minutes, three different wins. Pick yours.
How it helps you
Understand LLMs, RAG, MCP and agents well enough to explain them in an interview, not just name them.
Turn notes and PDFs into a cited study notebook that refuses to make things up.
Leave with a working Python agent you can put on GitHub and talk through line by line.
Know which AI tools are free, which are safe, and when not to trust the answer.
You walk out with
A project for your résumé: your own AI agent
A first SKILL.md file and a reusable study brief
A certificate of participation
Goal: say "I built one", not "I used one".
How it helps you
Stop rewriting the same prompt: set up a standing brief that writes your weekly status email.
Ask questions of your own documents and get answers with citations you can check.
Connect AI to Drive or Calendar safely, and know the three things never to connect.
Build a small agent that reads your notes, totals the numbers and drafts the update, then waits for your approval.
You walk out with
An agent you edit for one real weekly task
The never-paste card for company data
A one-week plan to use it on Monday
Goal: one weekly task off your plate.
How it helps you
Get the real vocabulary of AI work: tokens, context windows, system prompts, RAG, MCP, tool calls, evals.
See how a production engineer thinks about models, cost and safety, not just demos.
Build and evaluate an agent yourself, so your first AI project is behind you.
Leave with a clear path for what to learn next and where to practise for free.
You walk out with
Your first working AI agent and its eval card
A map of the AI stack, from model to agent
A daily series to keep learning: 100 Days of GenAI
Goal: your first AI project, done in a morning.
What you build in the room
Seven things. All yours, all working.
01
Lab 0 · 3 min
A locked-down account
Your ChatGPT, Gemini and Claude privacy settings, set on purpose, plus the never-paste card.
Use it for: knowing exactly what you share before you share it.
02
Lab 1 · 5 min
The three-step check
Catch a model guessing, ask for its source, and verify outside the tool.
Use it for: never forwarding a confident wrong answer again.
03
Lab 2 · 8 min
A standing brief + your first skill
A Claude Project with your instructions and a real document, plus a five-line SKILL.md for one job you do weekly.
Use it for: one-line requests that come back in your format.
04
Lab 3 · 6 min
A grounded notebook
NotebookLM on your own PDF: answers with citations you click, and a refusal when the answer isn't there.
Use it for: reports, syllabi and policies you can actually trust.
05
Lab 4a · 4 min
A model shortlist
Two models picked for your own task from the leaderboards the industry uses, with free-tier and privacy notes.
Use it for: choosing a model on evidence, not hype.
06
Lab 4b · 7 min
A connected tool, safely
Drive or Calendar plugged into Claude through MCP, the approval prompt you read, and the lethal-trifecta card.
Use it for: letting AI act on your files without leaking them.
07
Lab 6 · 18 min
Your own AI agent
About 60 lines of plain Python in Google Colab on a free model. It reads a file, calculates, drafts, and asks you before sending.
Use it for: automating one real task, and proving you can build.
The full curriculum
Five minutes of why. Then your hands on the keyboard.
No slide marathons. Every cycle is a short, sharp explanation, a live demo, and then a lab where you build the thing yourself while I walk the room. You leave each cycle with something that works, not notes about something that might.
51of 90 minutes are yours to build7labs6cycles0prior coding
Tap any cycle to see exactly what's taught and what you build.
Open & lock your settingsWhere AI helps and where it hurts
+
Topics
The job question, with real numbers
The jagged frontier of AI ability
What gets stored, and how to control it
Lab 0 · 3 minSet the privacy toggles in ChatGPT, Gemini and Claude. Read the never-paste card.
6–17
Inside the LLMTokens, transformers and attention
+
Topics
How an LLM answers, in four steps
The transformer and attention, in plain English
Who makes the models: GPT, Claude, Gemini, Llama, Qwen, DeepSeek
Why it's not a database, and the three-step check
Lab 1 · 5 minCatch a model guessing, ask for its source, write down what it did when it didn't know.
17–31
Prompts, standing instructions & skillsWrite it once, reuse it forever
+
Topics
How a prompt should be: Role, Task, Context, Format
System prompts, Projects and Gems
What "memory" really is
Agent Skills: the SKILL.md open standard
Lab 2 · 8 minBuild a Claude Project that writes your status email, then write your first five-line skill.
31–42
RAG by doingThe open-book exam
+
Topics
Training cutoff vs your documents
"In my sources" vs "from training"
Embeddings in one line
Web search vs deep research
Lab 3 · 6 minGround NotebookLM in your PDF, click a citation, and earn a refusal.
42–51
Models & finding modelsPick on evidence, not hype
+
Topics
Model families, sizes and cost per token
Open weights vs hosted models
OpenRouter Rankings, Arena, Artificial Analysis
Free tiers for building, and their privacy terms
Lab 4a · 4 minShortlist two models for one task you do, and pick the account you'd use at work.
51–64
Tools, MCP & the agent loopA tool call is a request, not an action
+
Topics
Think, act, observe, repeat
Workflow vs agent: who decides the next step?
MCP: the USB-C port for AI
Prompt injection and the lethal trifecta
Lab 4b · 7 minConnect Drive or Calendar to Claude, read the approval prompt, fill the safety card.
64–86
Build your agentThe 20-minute build
+
Topics
The 60-line agent, read top to bottom
System prompt, tools, loop, stop, approval
What an eval is, and why answers vary
Lab 6 · 18 minRun it in Colab on a free model, change the task to your own job, run it three times and score it.
86–90
CloseWhat you take home
+
Topics
Exit check: three questions
The 4D frame: Delegation, Description, Discernment, Diligence
What to learn next
Take homeKit, script, Colab steps, slides and certificate by email.
The 20-minute build
60 lines. 4 parts. An AI that works for you.
This is the real code you'll run on Sunday. No framework, no black box: a brief, two tools, a loop and one line that makes it ask before it acts. By the end you'll know what every line does, and you'll have changed it for your own job.
1line to swap the model2tools it can use8steps max, never runs away0emails sent without your yes
# 1 · one client, any provider
client = OpenAI(base_url=BASE_URL, api_key=API_KEY)
# 2 · the standing brief (your Lab 2 skill goes here)
SYSTEM = "You are a careful office assistant. Use tools
instead of guessing. Ask before sending."# 3 · tools are plain Python functionsdefread_file(path): ...
defcalculate(expression): ...
NEEDS_APPROVAL = {"send_email"}
# 4 · the loop: think, act, observe, repeatfor step in range(MAX_STEPS):
reply = client.chat.completions.create(...)
if not reply.tool_calls: return reply.content
if name in NEEDS_APPROVAL: ask the human
Part 1
Client & model
One line connects to OpenRouter, Gemini or Groq. Swap one line, swap the model.
Part 2
System prompt
Your standing brief, in code. Your skill's five steps go here.
Part 3
Tools
Functions the model may ask for. Sending sits in the approval set.
Part 4
The loop & the stop
Run tools until the model answers in plain text, with a hard limit, and a human before anything leaves.
Everything runs on free tools
₹0. Zero installs. Your browser is the whole lab.
No credit card, no downloads, no "works on my machine". If your laptop can open Chrome, it can run every lab, including the Python agent.
FREEClaudeProjects & connectors
FREEChatGPTprivacy controls
FREEGeminiactivity settings
FREENotebookLMgrounded answers
FREEGoogle ColabPython in a browser
FREEOpenRouterfree models, no card
₹0to attend, and to run every lab
0software to install
1browser tab per lab
What you take home
A kit you'll actually open again.
Pre-session checklistAccounts, keys, togglesLab cardsEvery step, with fallbacksThe agent scriptExplained in plain EnglishOne-week planPut it to work on Monday
Plus the slides and a certificate of participation, sent by email after the session.
Your instructor
Lakshmanan, AI engineer, now building agents.
I build production AI agents. About eight years in AI: I've built models, fine-tuned them for specific jobs, and now I build agents. In the last five months I built ten different AI products, including agentic AI projects. I teach from what I've actually shipped: how agents work, why they break with real users, and what's new in AI minus the hype.
~8 years in AIModels, data, production
Fine-tuningModels trained for specific jobs
AI agentsTools, loops, MCP, approvals
10 AI products in 5 monthsIncluding agentic projects
I post one short lesson a day as 100 Days of GenAI: how LLMs work, how agents are built, and what survives real users.
The labs have been tested end to end on free accounts, including the agent. Real attendee quotes will appear here after the session, not before. Come and be one of them.
No. You paste a tested 60-line Python script into Google Colab and change one line. Every part is explained in plain English, and the kit walks you through each click.
Is it really free?
Yes. The session is free, and every lab runs on free accounts: Claude, ChatGPT, Gemini, NotebookLM, Google Colab and a free OpenRouter key. No credit card anywhere.
What do I need on the day?
A laptop (not a phone) with Chrome, a stable connection, the free accounts in the checklist, and two PDFs that aren't confidential. The checklist arrives two days before. It takes about 20 minutes.
Will there be a recording?
No. It's a live, hands-on session: the value is in what you build. You keep the kit, the script and the slides.
I already use ChatGPT a lot. Is this too basic?
If you've never written a system prompt, grounded a model in your own PDF, connected a tool through MCP or built an agent loop, there's a lot here. The last 20 minutes is a real Python build.
Is it safe to use with work data?
We use sample data only. You'll learn exactly what not to paste, how free tiers use your data, and which setups are risky, so you can use AI at work responsibly.
How do I get the Zoom link?
Register on Luma, then join the WhatsApp channel. The link is posted there and emailed to registered guests on the morning of the session.
Sunday 25 October · 10:00 AM IST
Your seat is free. The agent is yours.
90 minutes, live, on your own laptop. Leave with seven things you built, including an AI agent.
01
02
03
04
05
06
07
