From Software Engineer to AI Engineer
You already know how to build software with AI. Now learn to build AI into the software you ship.
LLMs are becoming part of real products. Learning to design, test and integrate those capabilities can help you stay relevant as the industry changes.
Across five in-person sessions, build one AI-enabled product that uses an LLM, draws on domain knowledge, calls tools and manages its context as the task grows.
This cohort is closed now
Please register interest so that we can let you know when there is a future cohort available.
Stay relevant in today's market.
Build an AI-enabled product you can put on your CV and show on GitHub.
You'll build a working AI-enabled product during the course. Add it to your CV and show the code on GitHub as a real-world portfolio project: something you can walk through with a hiring team, from the engineering decisions to the tests.
Build on the skills you already have.
You've built products. This is your chance to engineer the AI capabilities inside one.
Bring your experience with APIs, data, tests and product code. Go deeper on evals, retrieval, tools, context and security by applying them to your own project—not just prompting an AI coding assistant.
AI tools can help you write code during the course. The skill you're developing is deciding what the AI feature should do, how it fits into a product, and how to tell when it fails.
What you'll need
- A laptop with Python and your own AI coding tool
- Time for 5–8h of homework a week
Not for: complete beginners, or anyone after ML theory or model training. Software engineering is one route into this course; see the general AI Engineering course page for the broader audience.
Five Sundays, one product.
Each session adds a capability to your project. Build your own implementation alongside other students, then continue it as homework.
- 10:00–11:00Review homework and Q&A
- 11:00–12:00Introduce the session's topics, with coding examples
- 12:00–12:30Lunch break, food provided
- 12:30–16:00Work on your project
- Five Sundays, 10:00–16:00, in person (1, 8, 15, 22 and 29 November 2026)
- Lean on AI tools for coding so you can focus on the engineering decisions
- Missing one session has an impact, but there are ways to deal with it; missing more than one is not recommended, as it may significantly impact the outcome
This cohort is closed now
Please register interest so that we can let you know when there is a future cohort available.
Learn from someone who's made the move.
A software developer with 25 years of experience who has already transitioned into AI engineering.
Foundations
Start where engineers are strongest: define the product, make a working baseline and test its behaviour before adding complexity.
Project
- Define your project, its expectations and what is essential
- Set up the tools you'll use throughout the course
- Protect against untrusted user input
LLM API
- Understand the available LLM APIs
- Make calls and build up messages from a conversation
Testing
- Use an eval suite as regression tests for the AI feature
- Optimise prompts using automated tests
Retrieval (RAG)
Move beyond a model's built-in knowledge: add domain documents and test whether the right information reaches the answer.
Core
- Add a document with domain knowledge using embeddings
- Use a text embeddings model to create the embeddings
- Store and query embeddings in PostgreSQL pgvector
- Prompt the LLM to give references for its answers
Testing
- Retrieval regression tests (query → expected chunk)
- End-to-end golden set for the feature
Security
- Indirect prompt injection via retrieved documents
- PII in the document corpus
Extras
- Chunking strategies, hybrid search, reranking and failure modes
- When not to use RAG
This session is deliberately light, leaving time to catch up if week 1 was a difficult start.
MCPs & Tools
Give the AI feature access to product capabilities while retaining the engineering controls you'd expect around any integration.
Core
- Understand tool calling
- Build ReAct loops with tool calling
- Create a custom MCP for your tools
Testing
- Assert tool calls using mocked tools
- Test when the loop terminates
Security
- Least-privilege tools and argument validation
- Confirmations for dangerous actions
- MCP trust boundaries
Extras
- Planning patterns and guardrails
- When agents are overkill
Context management
Treat context as a limited resource: decide what to retain, what to compact and how to spot failures when information is lost.
Core — compaction
- Context as a capped budget, triggers and keep-zones
- Summarisation vs. pruning and masking
- Context drift and dropped-constraint failure modes
Testing
- Constraint-preservation tests through compaction
- Run the full regression suite
Security
- Compaction must not drop safety constraints
Extras
- Rewind/undo: checkpointing, conversation rollback, regeneration, agent-state recovery
- Long-term memory: user-scoped cross-session memory; minimal in-house build, with Mem0 as an off-the-shelf shortcut
- Memory testing and security: recall tests, PII in memory, cross-user scoping, memory poisoning
Advanced topics + demo day
Step back from the implementation: consider cost, how specialised agents divide work, and what it takes to harden an AI feature. Then show what you built.
10:00–12:00 · Presentation
- Cost management: token accounting, caching economics, model routing, cost caps
- Distributed/specialised agents: their own context and specialty, context isolation, when to split
- Hardening checklist: rate limits, abuse, PII in logs/traces, key review
12:00–16:00 · Demo day
- Brownbag demos
- Completion feedback
Engineering by doing.
- Learn by doing, not by powerpoint
- Everyone moves at their own pace
- Students can choose what they want to focus on
This cohort is closed now
Please register interest so that we can let you know when there is a future cohort available.