Become an AI Builder

That is a fair sample of this course.

It is not a produced course. It is sixteen multi-hour Zoom classes, recorded as they happened, in which Sahil builds software with AI coding agents in front of a cohort of beginners — and debugs their broken installs while they watch. Fifty-five hours, which makes it by a wide margin the longest thing in the catalogue, and the least edited.

What it actually contains, session by session.

Day one is the only lecture: how these systems work, from pattern recognition through machine learning, embeddings, token prediction and why Hindi costs more per prompt, to fine-tuning against retrieval, temperature, hallucination, running models locally, and agents. Days two and three move to using AI as an assistant — the interfaces, deep research, skills, connectors, running an agent against a real inbox.

Days four and five are generative media and the first vibe-coding session. Days five through ten are one long build: a CRM, taken from a written spec to a working web app, then packaged as a desktop application, then shipped on GitHub, with a companion mobile app along the way. Day eleven is the one that matters most — stop shipping localhost. A full deployment on screen: domain, DNS records, environment variables, the secret that must never reach GitHub, and the site going live. The Google login fails on the first attempt and he leaves it broken.

Days twelve and thirteen move to the business: the model canvas worked against his own course, a news automation pipeline dictated aloud phrase by phrase, ad platforms wired into an agent, and — after the recording restarts — the clearest teaching block in the entire course, on pricing. Days fourteen and fifteen are production engineering: what scaling actually requires, retrieval done properly, cost and latency control, hosting from shared plans up to buying a second-hand server, and containers.

Day sixteen is three hours of him opening about thirty student projects in a browser and taking each one apart. That session is the most useful thing here to watch twice, because the criteria repeat: is it real or is it dummy data? He clicks save, refreshes, and checks whether anything persisted. Can he even open it. Does it log you out. Does it work on a phone. And the question he puts to almost every project that works — what's next, and where's the business?

Two honest warnings, because they change what you should expect.

One. The transcript this catalogue indexed is damaged enough that we will not repeat a single number, price, model version or benchmark from these sessions as fact. Where he states a limit or a version, watch him say it and check it yourself. What survives reliably is the shape of the work, the tool inventory, the sequence he does things in, and the reasoning.

Two. This is a recording of a live cohort, with everything that implies. Sessions start late. Twenty minutes of a class can be setup. Some builds do not finish — the ads integration hits an approval wall and stops, a deployment ends partially working with the landing page explicitly undone. He changes his recommendation mid-session and abandons tools on air. If you want a polished curriculum, this is the wrong purchase. If you want to watch somebody actually do this work at the speed it actually goes, there is nothing else like it in the catalogue.

Taught in Hinglish.