Generative AI Course — Build With LLMs, Not Just Learn About Them

The practical Python you need, then only the ML concepts that come up later: what a model actually is, what training versus inference means, what an embedding represents, and how tokens get billed.

We skip the rest. You will not spend three weeks on gradient descent for a job where you’ll never train a model from scratch. If you already write Python daily, this module is revision and you can treat it that way.

Transformers at the level you need to reason about behaviour, not to implement one. Tokens, context windows, temperature, what “the model forgot” actually means.

Then the practical half: choosing a model. Each of the big three has different strengths, different pricing per million tokens, and different latency characteristics. Most teams pick one because somebody read a benchmark on Twitter. You’ll learn to pick on cost and latency for your specific task, which is what an interviewer wants to hear.

Structured prompting, few-shot examples, chain-of-thought, output formatting, and evaluation.

The emphasis is on techniques that survive model updates. A prompt trick that worked on one model version and broke on the next isn’t engineering, it’s superstition. You’ll build a small evaluation harness so you can tell the difference when your prompts stop working.

Deeper reading: the six prompt engineering techniques that matter.

The module that gets the most time, because it’s what most GenAI jobs actually involve.

You build a document Q&A system end to end: chunking, embeddings, a vector database, retrieval, and grounding the answer so the model stops inventing things.

The part worth knowing in advance: most RAG systems fail at retrieval, not generation. The model isn’t the problem. Your chunks are too big, or too small, or split mid-sentence, or your embedding model doesn’t understand your domain vocabulary, or you’re retrieving five documents when the answer needed three specific paragraphs. We spend real time on evaluating retrieval quality, which is the thing most tutorials skip entirely and most interviewers ask about.

Background reading: what RAG actually is.

Chains, memory, tool calling, structured output, and error handling for a system that fails in ways normal software doesn’t.

The distinction that matters here is between a notebook demo and an application. A demo works once, on your machine, with your example. An application handles a model timeout, a malformed JSON response, a user asking something out of scope, and a rate limit at 4 PM on a Tuesday. We build the second kind.

How to adapt an open-source model to your domain using LoRA on affordable hardware. Dataset preparation, training, evaluation, and deployment of the result.

Equally important: when not to fine-tune. Most people who want to fine-tune should be improving their retrieval instead. Fine-tuning teaches a model style, format and domain vocabulary. It does not reliably teach it facts, and it doesn’t stop hallucination. Getting that distinction right in an interview separates you from most candidates immediately.

What agents add beyond a chat interface: tool use, multi-step workflows, and the planning loop.

Deliberately an introduction. Agents are a large enough subject to deserve their own programme, and we have one. This module gives you enough to understand what you’re looking at and to decide whether you want to go further.

See the Agentic AI Course, or read AI agents explained first.

FastAPI services, Docker containers, AWS deployment. And then the parts nobody teaches.

Token cost management, because a feature that costs ₹40 per user per day doesn’t ship. Caching, so you stop paying twice for the same question. Streaming, because a nine-second wait with no output feels broken even when it’s working. Rate limits and retry logic. Monitoring what your LLM is actually saying in production. And what you do when a provider deprecates the model your product is built on, which will happen.

This module is the one interviews test hardest, and it’s the one most GenAI courses treat as an afterthought. If deployment interests you more than modelling does, look at our DevOps & Cloud program too.