LLM Engineering & RAG
Eight weeks building retrieval-backed LLM systems that hold up on real documents and real questions. You will build a working RAG system, measure it honestly, and learn the three failure modes retrieval cannot fix.
Join the waitlistWho should take this program?
You'll get the most from this if
- Confident Python
- Comfortable calling APIs
- Basic Git
This is not for
- Complete beginners — start with Python for AI Engineering
- Anyone looking for prompt tips rather than systems work
What you will build
- A RAG system over a messy real document set
- A retrieval evaluation suite separating retrieval quality from answer quality
- A cost and latency report with the trade-offs you chose
What you learn, week by week
Module 1 — Prompting for reliability
- Get structured output every time
- Handle refusals and truncation as expected states
- Measure variance across runs
Module 2 — Chunking and ingestion
- Chunk documents so retrieval works
- Handle tables, PDFs and layout
- Keep provenance for citation
Module 3 — Embeddings and vector search
- Choose and evaluate an embedding model
- Build a pgvector index
- Understand what similarity is missing
Module 4 — Hybrid search and reranking
- Combine keyword and vector retrieval
- Rerank for precision
- Measure the gain honestly
Module 5 — Answer generation and citation
- Ground answers in retrieved context
- Cite sources reliably
- Refuse when context is insufficient
Module 6 — Evaluating RAG
- Separate retrieval metrics from generation metrics
- Build a golden set from real questions
- Run evaluation in CI
Module 7 — Cost, latency and caching
- Profile where time and money go
- Cache correctly
- Route between models
Module 8 — Where RAG breaks
- Recognise the three failure modes retrieval cannot fix
- Decide when to use an agent instead
- Present your trade-offs
How this program handles evaluation and governance
A confident wrong answer is the characteristic failure of retrieval systems, and it is invisible without measurement. Week 6 makes evaluation a build requirement: retrieval quality measured separately from answer quality, a golden set drawn from real questions, and citation you can check. Week 8 covers refusing to answer, which is often the correct behaviour and almost never the default.
Tools and stack you will use
Fees
Request pricing
Fees depend on cohort, format, and whether this runs for an individual or a team. We'll send the full breakdown.
Get fee detailsQuestions people ask
Do I need a vector database subscription?
No. We use PostgreSQL with pgvector, which you can run locally.
Is this enough to build production RAG?
It is enough to build it and to prove whether it works. Operating it at scale is covered in AI Systems in Production.
Does this cover fine-tuning?
Briefly, mostly to explain why retrieval usually beats it for the problems people reach for it on.