Portfolio 1
Enterprise RAG
Hybrid search and re-ranking over a large document set, plus an evaluation pipeline so you can raise accuracy on purpose.
Production Generative AI Systems
This course is designed for software engineers with intermediate AI experience who want to move beyond calling APIs. You will customise LLMs, build retrieval and agent systems you can evaluate, and ship them with latency, cost, and safety under control.
Created by Baljeet Dogra
Course Objectives
- Customise large language models with advanced prompting, retrieval-augmented generation, and parameter-efficient fine-tuning.
- Build retrieval and agent systems you can evaluate — with frozen test sets, citations, and traces — not just a happy-path demo.
- Choose foundation models on cost, latency, and quality, and write down a default, a fallback, and a hard “do not use”.
- Optimise serving with containers, latency budgets, timeouts, and a simpler path when the agent is on fire.
- Apply guardrails, privacy redaction, human review, and spend caps before real users depend on the system.
- Deliver an industry capstone: metrics written first, the thinnest path that hits them, and a 12-minute demo you can defend.
What you will build
Five systems you can demonstrate, defend, and extend. Each lab has a brief, acceptance checks, and a write-up.
Portfolio 2
Fine-tuned LLM API
Open model adapted with LoRA or QLoRA on domain data, compared to a baseline, and exposed behind an API.
Portfolio 3
Multi-agent system
Agents that plan, call tools, and coordinate a multi-step workflow. Runnable end to end, not a slide of boxes.
Portfolio 4
Deployed GenAI API
Containerised serving with baselined latency, logging, and a health path you would actually put behind a load balancer.
Capstone
Industry system
Pick a real problem in healthcare, BFSI, manufacturing, or your workplace. Design the architecture, implement retrieval and agents, add safeguards, measure it, and ship a working prototype with a technical report.
32-week curriculum
Five modules plus capstone. Expand a module for weekly labs. Content stays searchable when closed.
01 Foundations of GenAI & LLMs Weeks 1–6 · 18 hours live
You leave this module able to explain how transformers process language and to justify a foundation-model choice on cost, capability, and constraints.
1.1 Tiny language model
Train a toy “guess the next character” model on your laptop so loss, training, and failure are concrete — before we talk about ChatGPT-scale systems.
Lab
Load a short text file, train until the error number drops, generate a bit of text, and write half a page on what an API would hide compared with this toy.
1.2 Tokens and attention
See how a real model chops a sentence into pieces, and why a bad split later breaks document search.
Lab
Tokenise the same sentence two ways, look at one attention picture, and mark where a PDF table would be cut from its header.
1.3 Pick a model
Same 20 tasks on three models. Quality, speed, and cost — not a marketing winner.
Lab
Fill a spreadsheet, then write a one-page memo: default, fallback, and one hard “do not use”.
02 Advanced prompting & RAG systems Weeks 7–12 · 18 hours live
RAG is easy to demo and hard to deploy. You will design retrieval with measurable, improvable quality.
2.1 Let the model use tools
A loop that calls your functions, retries bad JSON, and refuses jailbreaks — not one lucky screenshot.
Lab
Two stub tools, 30 fake user turns, and traces for success, retry, and a clean refusal.
2.2 Search your documents
RAG: look up passages first, then answer. Compare three ways of slicing the same corpus.
Lab
Index ~200 pages, write 25 questions with known answers, and pick a winner with recall@5.
2.3 Better search + a score
Keyword search plus meaning search, a reranker, citations, and a test you rerun after every change.
Lab · Portfolio 1
Ship a measured RAG stack with a before/after eval table and a README a classmate can rerun.
03 LLM fine-tuning & alignment Weeks 13–18 · 18 hours live
The module most programmes rush. Three full weeks on PEFT because proprietary data is where leverage actually shows up.
3.1 Prompt, search, or train?
Three levers. Score the job. At least one brief must be “do not fine-tune”.
Lab
Three scored product briefs plus a one-page defence a budget holder would accept.
3.2 Light fine-tuning
LoRA/QLoRA: train a small add-on, not the whole model. Compare two sizes. Check it did not forget general chat.
Lab
Holdout eval table, GPU hours, and a written rank choice — even if the adapter loses.
3.3 Serve the tuned model
HTTP API, a switch back to the original model, smoke tests through the API.
Lab · Portfolio 2
README with curl examples, eval table, and a one-change rollback plan.
04 Multimodal & agentic AI Weeks 19–24 · 18 hours live
AI that sees, reasons, and acts. You will build multimodal pipelines and agent workflows for real business use cases.
4.1 Documents and pictures
Photos of pages → JSON fields, then a spec → labelled diagram. Measure the lies, don’t admire the pictures.
Lab
Ten redacted pages scored field-by-field, plus three hallucination types you actually saw.
4.2 Several agents
Planner, workers, supervisor. A log of a success and a recovered failure — not a chatbot with extra names.
Lab · Portfolio 3
A graph that runs locally, two saved traces, and a diagram a human can follow.
05 Deployment, optimisation & AI safety Weeks 25–32 · 24 hours live
The difference between a demo and a product is engineering decisions. You graduate able to take a business problem, pick the right LLM or agent approach, and deploy it safely.
5.1 Timeouts and fallbacks
How slow is too slow, stop waiting on dead tools, and a simpler search-only path when the agent is on fire.
Lab
An SLO, ~100 requests, a circuit breaker, a degraded response, and a one-page runbook.
5.2 Put it in a container
Docker, a health URL, latency numbers at a rate you actually ran.
Lab · Portfolio 4
A compose file a reviewer can start, plus p95, QPS, and a short note on real GPU serving vs your stub.
5.3 Safety, privacy, spend
Filters, hide emails, human queue, token cap, one-page governance.
Lab
Show a blocked request, a capped request, and before/after redaction.
5.4 Industry capstone
One domain, metrics written first, thinnest path that hits them, 12-minute demo.
Capstone
Running prototype, diagram, eval, operational notes, and a report of what you cut. Reuse earlier labs.
Who this is for
Software engineers
You already ship production systems. You want retrieval, evaluation, and deployment — not another “call the API” tutorial.
ML and data practitioners
You can train models. You need to turn experiments into GenAI features with measurable quality, cost, and safety.
Technical leads and architects
You decide the approach. You need to evaluate designs and lead implementation, not only delegate it.
Prerequisites
- Python functions and basic data structures
- Calling or building HTTP APIs
- Comfort reading technical documentation
- Basic maths and statistics behind common ML ideas
Not a fit if
- You want a first course in programming or ML
- You only need a management overview of AI
- You want a certificate without the labs
Start with Python for GenAI if the Python bar is not there yet. Choosing which architecture before you go deep? Take LLM Apps: Architecture by Use Case. Prefer to deploy in week 4 and every month after? Take Applied GenAI Engineering — same destination, ship-first route. Already shipping and need to decide under pressure? Take Scenarios as an overlay.
Questions
How does this relate to the LangChain and LangGraph courses?
Those tracks go deep on one stack. This course is the production path around them: model choice, PEFT, evaluation, serving, and safety. You will still write LangGraph-style graphs in module 4.
How does this relate to Applied GenAI Engineering?
Same destination — from calling APIs to a deployed RAG or agent system. This course is organised by topic (transformers, prompting, RAG, agents, then deployment). Applied GenAI Engineering is organised by shipped increments: first deploy in week 4, then every four weeks. Choose the route that matches how you actually need to learn.
Is this live or self-paced?
Live weekend teaching plus labs you complete during the week. Cohort dates are confirmed when you enrol. Recordings are for catch-up, not a substitute for the labs.
What compute do I need?
A laptop for prompting, RAG, and API work. Fine-tuning labs use cloud GPUs (QLoRA on modest instances). We specify instance types and estimated hours before that module so you can budget.
Ready to own the full path?
32 weeks from model internals to a deployed agentic system. Create an account to enrol in the next cohort.
Enrol now