02 / Programme
AI Engineering with Python
Build working AI systems in Python — agents, retrieval, and the production concerns that decide whether a prototype can leave the lab.
3 days · Up to 12 participants · On-site or remote · From £10,000
Who it is for
Software engineers and technical leads who will design, integrate and own LLM systems — not a room being shown a chatbot.
Who it is not
It is not an executive overview, and it is not a survey of products. Delegates write, test and run code.
What they will be able to do
- 01
Call, stream and constrain an LLM from Python, with prompts and structured outputs that can be tested.
- 02
Build a tool-calling agent with a registry, allowlists, rate limits, audit logs and an evaluation harness.
- 03
Stand up MCP servers and wire them into an agent over stdio, including live tools and structured data.
- 04
Index a corpus, retrieve with citations, and compare dense, sparse, hybrid and reranked search.
- 05
Extract facts into a graph, add memory with decay, and run ReAct and plan-and-execute loops.
- 06
Take a system toward production: tracing, retries, cost controls, guardrails, and a deployable FastAPI service.
Outline
Day 01
Build a working agent
- 00
Python fundamentals
The engineering substrate the rest of the week sits on: dataclasses and Protocols, CLI and logging, asyncio (tasks, queues, timeouts, cancellation), then HTTP with FastAPI and httpx.
- 01
Working with the LLM
Chat completions, message roles and parameters; token streaming; prompting for structured outputs, grounding and tool calling; a console chat with session storage.
- 02
Tool calling
Tool-call message format and agent state. A decorator-based registry (schema, validation, routing, errors). Safety rails: allowlists, rate limits, redaction, audit logs. An evaluation harness with golden tests, replay and deterministic mocks.
- 03
MCP servers
MCP as a protocol: tool discovery, schemas, calling conventions. FastMCP servers over stdio, dynamic discovery, and practical tools — structured data, web fetch, file I/O — connected to a live agent.
- 04
A research assistant
Day 1 closer. Model selection, structured outputs, multimodal (vision and transcription), FastAPI with SSE streaming, and MCP tools inside a web API. The agent leaves the console.
Day 02
Knowledge and retrieval
- 05
RAG fundamentals
Chunking (size, overlap, structure-aware), embeddings, vector stores. Dense, sparse, hybrid retrieval and reranking. Grounded prompting with citations, a comparison with and without retrieval, RAG evaluation, then the same index exposed as an MCP server.
- 06
Structured facts
Pydantic and JSON Schema structured outputs. Fact extraction with provenance and confidence. Knowledge graphs: entities, relations, path queries. Grounded QA that must cite [Fact N] or refuse.
- 07
Agent memory
Short-term session buffer versus long-term profile. Summarisation against a token budget. Decay and “do not remember”. Memory exposed as an MCP server so the agent can read and write it as a tool.
- 08
Structured workflows
Day 2 closer. ReAct (reason, act, observe), plan-and-execute with replanning, tool routing, and structured execution traces — workflows as software, not a prompt that hopes.
Day 03
Ship it
- 09
Multi-agent systems
When multiple agents help, and when they add failure modes. Roles (router, specialist, critic). Coordination: supervisor, swarm, debate, blackboard. Shared context, scoped tools, peer handoffs, consensus and conflict.
- 10
LangChain, as a choice
LCEL, prompt templates, output parsers, AgentExecutor, RetrievalQA, LangServe. The same loops already written by hand, now in a framework — so the room can judge the abstraction rather than start inside it.
- 11
Edge topics
Pick-and-choose, according to the room: hybrid search, reranking, HyDE, agentic RAG, citation verification, web-search backends, text-to-SQL, LLM evaluation, fine-tuning datasets, guardrails, semantic caching, multimodal RAG, contextual chunking.
- 12
Productionisation
Tracing, retries, circuit breakers, cost controls, deployment. Capstone: a production FastAPI backend for a real interface — RAG with hybrid search and reranking, agentic RAG with tools, or chat + MCP with cost controls.
How it is taught
Live engineering education. Delegates work in a Python lab with pytest on every exercise. Demos are run, not slid through. The programme can be tailored to the organisation’s stack; the spine stays the same.
Delivery and price
From £10,000 for up to 12 participants, on-site or remote. Talk to us to confirm dates, location, and any tailoring.
Instructor
Nicholas Johnson, AI architect and software engineer. Degree in Artificial Intelligence; around twenty years of professional technology training. About.