From scratch, first
Dot products, backprop, softmax, a minimal retriever, a tool-calling loop — built by hand in plain Python before any framework appears. The building is the learning. The framework is just the relief afterwards.
Most AI courses start where the abstraction starts. This one starts four layers underneath it — you write attention by hand in plain Python before you're allowed to type torch. Then you build the systems, ship them, and defend them in an interview.
One line. Same maths. The difference between an engineer who reaches for it and one who is trapped by it is whether they could have written the twenty lines on the left.
They're the reason this takes four months instead of four weekends.
Dot products, backprop, softmax, a minimal retriever, a tool-calling loop — built by hand in plain Python before any framework appears. The building is the learning. The framework is just the relief afterwards.
Attention Is All You Need. LoRA. RAG. ReAct. DDPM. Guided reading of the actual papers, mapped to the week you need them — so you learn to read a paper, not just a blog post about one.
Deployed, served over an API, cost-budgeted, monitored, evaluated. A notebook that only runs on your laptop is not an AI engineering portfolio. It's homework.
Live sessions, graded assignments with real code review, and debugging support between them. The depth is the point — so the timeline is honest rather than flattering.
Linear algebra, calculus and matrix calculus, probability, core ML — each concept paired with a lab where you build it from scratch and break it. Ends with a neural network written by hand in NumPy.
Transformers inside out, LLM training lifecycle, quantization and fine-tuning, RAG and vector search, agents and MCP, evals and guardrails, reasoning models, multimodal and diffusion — theory session, then a practical session that builds it.
Serving models over an API, deployment and infra, production concerns (cost, latency, caching, observability). Then portfolio polish, an AI-engineering resume, and a live mock interview.
Total ≈ 15–16 weeks. Depth, a fresh-graduate starting point, and speed cannot all be maximised. I protect depth. The cost is time.
Just you and me. The pace bends to you, the capstone is scoped around the roles you're actually targeting, and nothing gets skipped because someone else in the room already knew it.
A live group, kept deliberately small — small enough that I still review your code myself and still know your name. You get the room, the questions other people ask, and people to build alongside.
I teach both myself. If you're not sure which fits, that's the first thing we'll work out on the call.
Built in this order, each one leaning on the last. This is what you actually walk into an interview with.
MLP and backprop by hand in NumPy, then a trained MNIST classifier.
QKV, multi-head attention, causal masking — coded from the paper, not copied from a repo.
LoRA / QLoRA: dataset prep, training run, evaluation, and a write-up of what changed.
Chunking, embeddings, a vector DB, reranking — live on the internet, not on localhost.
A customer-support agent with real tool use, plus a multi-agent system wired through MCP with memory.
LLM-as-a-judge, guardrails, and the ability to argue about metric validity and bias.
Scoped, architected, built, evaluated, deployed, demoed on video. The one you'll be asked about.
You will find programmes that promise more than this. Read the right-hand column and ask yourself how they can possibly keep those promises.
Everyone does. It's the second thing anyone wants to know and the first thing most sites hide behind a form. Here's mine.
Any figure I put on this page would be a guess. I don't know yet whether you're starting from a blank page or halfway there, or whether you want me one-to-one or in a room with others. Those change the answer, so I'd rather ask than assume.
So: get on a call, tell me honestly where you are, and you'll have a real number inside ten minutes. No forty-minute pitch first. If it's out of reach, say so and I'll tell you what I'd do in your position instead. And if I don't think this is the right thing for you, I'll say that too — I'd rather lose the money than take it from someone I can't help.
I'm a software engineer with four years of building things that had to work in production. Before this I spent a long time being the person other people came to when the maths stopped making sense — around 1,300 students, at last count. I sat GATE Computer Science and came out at All-India Rank 720.
I'm not an Master of AI and I won't pretend to be one. What I am is an engineer who went and did the work — derived it, built it, broke it, shipped it — and who is unusually good at getting someone else to the same place. There's no team behind me, no ghost-written slides, no teaching assistant taking the sessions I said I'd take. You get me, every session, until you're done.
No forms into a void, no sales team. This goes straight to me, and I reply myself. If you're a fit we'll do a proper call; if you're not, I'll say so.