A software factory, built as a course

The factory: a ticket comes in from an issue tracker, a support widget or an alert. An agent takes it, writes the fix and opens a pull request. In the most aggressive setups it merges and deploys with no human in between.

Companies are building toward that right now. The question is not whether it works on a demo. It is what makes such a line safe to run every day:

• a harness that denies rather than suggests,
• evals you can trust more than your own eye,
• a cost per task, not per month,
• an automated reviewer that says what blocks prod,
• and a record of all of it.

This course builds those parts one module at a time. You pass a module by making its tests go green. Lessons carry the date their sources were checked and the versions they ran against. Where a claim needs a model, the course ran it once and committed the numbers.

Capstone reviewer, run on 12 pull requestsfixtures/live_review_run.json · claude-opus-5-5

0.80blocker recall

0.58precision of findings

10 / 12verdicts agree with the key

$0.70total cost, 131 s

This is what you ship at the end: a read-only agent that reads a PR against the repo's rules, runs the tests on the patched tree, ranks the blockers, and refuses to be talked out of it. The numbers are from one committed run, not a benchmark.

M0 · merged

Inference economics

Tool loop, streaming, prefix caching, cost ledger, batch builder

stages

7 stages

live record

synthetic transports

1. Tokens, prefill, decode
2. KV cache and batching
3. Prompt caching as prefix engineering
4. Cost levers: batch, effort, budgets, model choice
5. Structured outputs and strict tools
6. Context management: compaction, editing, memory
7. Interview angle: inference economics

Module	What you build	Stages	Live record	Status
M0	Inference economics Tool loop, streaming, prefix caching, cost ledger, batch builder merged	7	synthetic transports	merged
M1	Enforced harness Policy engine, PreToolUse hooks, durable sessions, caps, reviewer subagent merged	8	not recorded	merged
M2	Evals as engineering Trace loader, labels and kappa, schema-bound judge, trust and regression gates merged	8	judge 20/20 outcome	merged
M3	Observability and cost GenAI spans, cost from spans, budget guard, Langfuse export, router merged	8	9 spans · Langfuse 4.63	merged
M4	Retrieval as a component Chunker, pgvector store, hybrid RRF, reranker, metrics, cited answers merged	9	recall@10 0.78 · MRR 0.65	merged
M5	Security and compliance Injection detector, provenance and taint, secrets, remote MCP auth, audit log merged	9	injection denied · checklist 10/10	merged
M6	Serving and platform Load generator, local models, K8s manifests, autoscaling, gateway merged	8	gateway 6/6 routed · $0.057	merged
M7	Capstone: pre-review Rule loader, read-only reviewer, verify gate, blocker ranking, metrics merged	8	12 PRs · recall 0.80	merged

Order to take them: M0 → M1 → M2 → M3 → M4 → M5 → M6 → M7. Stages marked live need a Claude Code login; some need Docker, Go, Node or Ollama. miubais doctor tells you which.

How a module goes

1. uv run miubais start m0-cost-lab copies the starter into your workspace.
2. Read the lesson, write the code, run miubais check. Deterministic tests decide; the stage table on the site updates.
3. Stuck? miubais hint explains what the failing test checks. The tutor sees your diff and the test output, never the reference solution, and a filter drops anything that leaks it.
4. One stage per module is a judgment call: read the traces, write the taxonomy or the plan. Tests check structure and agreement with a key, not prose.

Start hereRead the capstone module first