AI Engineering Lab · 19-episode course
SaaS Support Copilot is a completed build-along course. Step by step, you create an AI assistant that answers usage, bug, and feature-feasibility questions from docs, code, Git history, the database, and logs—with citations, role-based access, and honest refusals.
$ course --status
▸ format build-along, one tested change per episode
▸ episodes 19
▸ languages EN · FR · ZH
▸ status complete
● read → cite → refuseCourse syllabus
SaaS Support Copilot is a self-paced build-along. Across 19 episodes plus one method lesson, you grow a copilot for a fictional ticketing SaaS—from the first typed data model to a capstone demo against a real model.
Course mission
Build a read-only copilot that answers usage, bug, and feature-feasibility questions about a SaaS application by reading its help docs, source code, Git history, database, and logs. Answers cite their sources, respect the caller's role, and refuse anything the evidence does not support. The copilot never performs actions; the course teaches the engineering around that boundary.
How the course works
Recommended order
Each episode builds on the previous one, so 0 → 18 is the easiest first pass. “Prompting this build” is a companion to the whole course: read it once you have a few episodes behind you.
Phase 1 · Episodes 0–3
Frame the problem, set up typed Python foundations, make the smallest LLM call, and turn model output into structured data.
Phase 2 · Episodes 4–7
Classify each question, give the agent safe read-only tools, drive a bounded loop, and keep conversation state.
Phase 3 · Episodes 8–13
Build RAG and a call graph from scratch, enforce permissions in retrieval, plan feasibility answers, refuse commands, and defend against prompt injection.
Phase 4 · Episodes 14–18
Add the HTTP boundary, evaluation and observability, local deployment, a production-readiness review, and a capstone demonstration.
What you will be able to do
Episode library
Every episode opens on GitHub, next to the code it builds, in English.
00 · Foundations
Model versus application versus agent, workflow versus agent, and a tour of the Loopline sample app.
Read episode01 · Foundations
Typed data models and tests, so every later episode hands data around without guessing.
Read episode02 · Foundations
A provider-neutral message interface with a fake client, built before touching a real API key.
Read episode03 · Foundations
Turn free-text model output into typed, validated values, with a retry on malformed output.
Read episode04 · Routing, tools & the loop
Decide what kind of question it is first—usage, bug, feature, or ambiguous—then route it.
Read episode05 · Routing, tools & the loop
Read-only tools over docs, source, Git history, the database, and logs, called through a registry with timeouts.
Read episode06 · Routing, tools & the loop
A loop the model drives—observe, decide, act—choosing one step at a time between calling a tool and answering.
Read episode07 · Routing, tools & the loop
Bounded conversation context and per-session memory, without confusing “this chat” with “this user”.
Read episode08 · Evidence & safety
Chunking, BM25, embeddings, and combining rankings—built from scratch so RAG stops being a black box.
Read episode09 · Evidence & safety
A call graph that answers “what would changing this function break?”—a question text search can't.
Read episode10 · Evidence & safety
Make the caller's role a first-class input to retrieval, enforced before the model sees any document.
Read episode11 · Evidence & safety
A plan-first workflow for feasibility questions: decide what evidence would settle it before spending a tool call.
Read episode12 · Evidence & safety
Recognize command-shaped requests and refuse them honestly, with zero model calls.
Read episode13 · Evidence & safety
Treat retrieved text as data, not instructions, and see an injected command blocked twice over.
Read episode14 · Ship & prove it
An HTTP boundary with validated requests, shared resources, request IDs, and clean errors instead of tracebacks.
Read episode15 · Ship & prove it
A golden question suite run against a real model, plus latency, token usage, and structured logs to catch regressions.
Read episode16 · Ship & prove it
Docker Compose with MySQL, secrets kept out of source control, and a standalone MCP server for database access.
Read episode17 · Ship & prove it
Draw the line between what the prototype has and what production needs; close one real gap, document the rest.
Read episode18 · Ship & prove it
One running system and a real model: the same question answered or refused depending on one header's role.
Read episodeBonus · Method
How an AI coding agent wrote every line: one focused, tested change per prompt, verified before it is written down.
Read episodeCurrent status
Completed: 19 episodes · MIT-licensed code. Everything runs locally on fictional sample data. Loopline is invented for this course and does not represent any real company's product, code, or data. This is a learning prototype, not a production-security guarantee—Episode 17 lists exactly what a real deployment still needs.