Generative AI end to end
Take this if you use generative models daily and want to understand and build the machinery underneath them.
What you’ll be able to do
You can explain and implement the main families of generative model, build a retrieval system and evaluate it honestly, and reason about the cost and latency of serving a model.
What it assumes
The Deep Learning phase or equivalent: you should have trained a transformer or a CNN yourself and be comfortable in PyTorch.
- Price
- ₹1,499
- Length
- 14 weeks
- Sessions
- 14
- Effort
- ~8 h/week · 112 h total
- For
- Class 11–12 · College · Working professional
Adds it to My courses on your dashboard so you can find it again. Payment is not open yet — nothing is charged.
What’s taught, in order
- 01
Autoencoders and latent space
Compression as the entry point to generation, then variational autoencoders and what makes a latent space samplable. The cleanest possible introduction to generative modelling.
- 02
Adversarial and diffusion models
GANs and their instability, then diffusion from the forward noising process to modern samplers. Two approaches to the same goal makes the trade-offs visible.
- 03
LLM internals and alignment
What is actually inside a served language model, then RLHF and alignment — including the honest account of what alignment does and does not guarantee.
- 04
Prompting, reasoning and evaluation
Prompting treated as an engineering practice with test sets and measured regressions, not a collection of tricks.
- 05
Embeddings, RAG and its evaluation
Vector stores, retrieval-augmented generation, and the evaluation strand that proves a system refuses to answer what it cannot support.
- 06
Multimodal models and inference cost
Models that span text and images, then quantisation, batching, caching and the arithmetic of what serving actually costs. Ends in a capstone.
What you build
Ship one generative system end to end with a real evaluation harness: pick a task, build the system, write an evaluation set that includes cases it should refuse or decline, measure quality and cost per request, then produce a written trade-off analysis of one architectural decision you made and the alternative you rejected.
Check it yourself against this
- The system runs on inputs it has never seen, from a clean start.
- The evaluation set includes negative cases, and results on them are reported.
- Cost and latency per request are measured, not estimated.
- One rejected alternative is described with the measurement that ruled it out.
- A failure gallery shows the system's worst outputs, not only its best.
- You state a use for which this system should not be trusted.
Then answer this
Diffusion models and language models generate in completely different ways — what do their training objectives nonetheless have in common?
Two or three sentences, in your own words. If you cannot, the course is not finished — go back to the module it came from.
What this course does not do
Nothing here is trained at frontier scale, and the course is explicit that model quality at this scale is a compute result you cannot reproduce. Video generation, speech synthesis and the legal and licensing questions around training data are all outside it.
Where this leads
The Generative AI phase page sets out all fourteen weeks with the model family covered in each, so you can see how autoencoders lead to diffusion and where the retrieval strand begins.
FAQ
- Will I train my own large language model?
- No, and no course honestly can — pretraining a competitive model costs millions in compute. You will build a small one from scratch, understand precisely what scaling adds, and spend the rest of the phase on the parts that are genuinely reproducible.
- How much of this is prompting?
- About one week, and it is taught as measurement rather than technique. The rest is mechanism, because prompting advice ages in months while an understanding of sampling, embeddings and retrieval does not.
- Does this cover agents?
- Only where retrieval touches them. Tool use, planning, memory, multi-agent systems and agent safety are the Agentic AI phase, which follows this one.
Other courses
This is a standalone course. It does not add weeks to the 78-week curriculum or change your roadmap. If you want the whole path instead, the full programme is ₹399/month · ₹2,000/year — see what it covers.