Stop watching AI tutorials. Start training models.
Eight weeks on real GPU hardware. You will quantise a model, fine-tune one on your own dataset, build a retrieval system that actually retrieves the right thing, put an agent behind guardrails, and serve the result as a production API with monitoring and a cost model. This is the program that requires hardware, and hardware is the reason it is small.
- Duration
- 8 weeks · around 64 hours
- Mode
- Live online, plus a hands-on GPU lab
- Cohort size
- 15–20 — capped by GPU capacity, not by marketing
- Hardware
- Your own sandboxed environment on Codnov GPU infrastructure
- Recordings
- Every session recorded, yours to keep
- Cohort 1
- Opens 17 October 2026 · dates sent to registered interest
Who this is for
- Working software engineers adding AI to what they already build
- Final-year computer science students who want production skill, not coursework
- Backend and data engineers moving toward ML and AI systems
- Technical founders who intend to build the AI part themselves
- Anyone who has used an AI API and hit the wall where prompting stops being enough
Prerequisites
- Comfortable with Python — functions, classes, virtual environments, pip
- Comfortable on the command line — navigating a shell, running scripts, reading a stack trace
- Basic Git — clone, branch, commit, push
- Helpful but not required: any prior exposure to PyTorch, Docker, or calling an LLM API
We hold to these. The cohort is small, and one person starting from zero slows everyone. Not there yet? Take AI Foundations first — its material is compressed into weeks 1–2 here, so you will be on solid ground.
Week by week
Foundations, compressed, and models on your own machine
The Foundations material at engineering pace: tokens, embeddings, transformers, training versus inference, the model landscape. Then a Python and PyTorch refresher aimed at the specific things you will need, and you get models running locally — loading weights, running inference, seeing where the memory goes.
You leave with: A local environment that runs open-weight models, and the vocabulary for the rest.
The GPU lab opens
The week the program becomes different from everything else. What a GPU is actually doing, and why VRAM — not compute — is usually the thing that stops you. Quantisation: making a model small enough to run without wrecking its quality, and measuring what you lost. Batching, and the trade-off between throughput and latency that decides what your system costs to run.
You leave with: The ability to look at a model and a GPU and know whether they fit.
Fine-tuning: your model, your data
You fine-tune a model on a real dataset. Not a demo notebook — LoRA and QLoRA, applied to data you bring or choose, with evaluation before and after so you can prove it improved. We cover when fine-tuning is the right answer and, just as importantly, when retrieval or a better prompt would have done the job for a fraction of the cost.
You leave with: A model you fine-tuned yourself, and evidence of what changed.
RAG, properly
Most retrieval systems fail quietly: they return something plausible that is not the right thing. This week is about why. Chunking strategies and how badly the wrong one hurts. Embedding models and how to choose one. Vector stores and what they actually do. Then the part usually skipped entirely — measuring retrieval quality, so you know whether your system is finding the right documents rather than merely finding documents.
You leave with: A retrieval system you can measure, and the ability to say why it is or is not working.
Agents, tools and guardrails
Letting a model take actions. Tool use and function calling. MCP for connecting models to real systems. Multi-step orchestration, and the failure modes that arrive with it — loops, runaway cost, confidently wrong actions. Then guardrails: constraining what an agent may do, and how to fail safely.
You leave with: A working agent with real limits on it.
Serving, deploying and paying for it
Turning what you have built into something other people can rely on. Serving with vLLM and Ollama. Designing an API around a model. Latency, and where it actually goes. Monitoring — knowing your system is degrading before a user tells you. And cost control, because an AI feature with no cost model is a bill waiting to arrive.
You leave with: Your work deployed behind an API, monitored.
The capstone
Build, deploy, demo, review. You ship your project, present it, and get a structured review against the published rubric from a Codnov engineer. This is the week the guarantee is about.
You leave with: A deployed, reviewed AI project, and your Codnov Certificate of Completion.
What you walk away with
- A deployed AI project, reviewed against a published rubric
- A model you fine-tuned yourself, with before-and-after evaluation
- A retrieval system with measured quality, not assumed quality
- Practical GPU skill: quantisation, VRAM budgeting, throughput and latency trade-offs
- A serving setup with monitoring and a cost model you understand
- Every session recording, all materials, and your starter repositories
- Codnov Certificate of Completion, with a unique ID and a public verification link
Sample projects
- A production knowledge assistant — RAG over a real document set, with retrieval evaluation
- A domain-tuned small model — fine-tuned for one narrow task, benchmarked against a general model
- A document-extraction pipeline — unstructured documents in, validated structured data out
- A voice agent — speech to text, a language model, text to speech, end to end
- A support triage system — classify, route, draft, and escalate when unsure
- A code-review assistant that reads a diff and comments usefully
- A multi-source research agent — plans, retrieves, synthesises, cites
- A self-hosted model serving stack — quantised, batched, monitored, with a cost model
Bring your own idea, or take one of these. Each comes with a prepared dataset and a starter repository — a project that is too ambitious is the most common reason people do not finish.
You finish with a deployed, reviewed AI project, or we re-run you free.
Requires minimum 75% attendance and on-time milestone submission. Full conditions at codnov.ai/training/terms.
Taught by Codnov engineers who build and run AI systems in production every day.
₹29,999+ 18% GST
₹35,399 including 18% GST
₹24,999+ 18% GST · early bird
₹29,499 including 18% GST
Early-bird price applies to cohort 1. Two-instalment payment plan available.
Nothing is paid on this website. Register your interest and we will confirm your seat with you directly.
Register your interest
We will come back to you with the full syllabus, the schedule, and whether there is a seat in the next cohort.
Codnov Certificate of Completion. Not a degree or diploma recognised under the UGC Act, 1956 or the AICTE Act, 1987. All prices exclusive of 18% GST. Project completion guarantee subject to minimum 75% attendance and on-time milestone submission; full terms at codnov.ai/training/terms. Open to participants aged 16 and above.