Capstone: One Real System
A business problem, end to end, defended live.
Everything above is components. The capstone is the proof that you can assemble them into something a company would actually run, and then stand in front of people and defend the trade-offs you made.
What you learn
- Scoping a real problem: support triage, document intelligence, or an internal copilot
- Assembling the full stack: control plane, retrieval, bounded agents, evals, gateway
- Deploying to your chosen target: AWS, GCP, Azure, on-prem, or agnostic
- Instrumenting cost and quality so you can answer questions with numbers
- The live defence: architecture, constraints, failure modes, cost, measurement
Tools you will actually touch
The assignment pack
Pick a real business problem, ideally from your own company or a company you want to work for. Build and deploy the system end to end, then defend it in front of a panel that will probe your constraints, your evals, and your cost model.
- Deployed and reachable, with the endpoint documented
- Control plane enforcing what the model may and may not do
- Retrieval or agent behaviour measured against a golden set
- Cost per request instrumented and attributable
- Live defence completed: architecture, failure modes, and measurements explained
A demo that works on the happy path. Cannot answer what it costs or how it fails.
Deployed, bounded, measured, and defended with real numbers.
The candidate anticipates the panel: knows the weakest part of their own system, has measured it, and has a credible plan for it.
The tutorial pack
The capstone playbook
- 1Week 16: scope ruthlessly, write the one-page architecture and the measurement plan
- 2Week 17: build the vertical slice end to end before broadening anything
- 3Week 18: instrument cost and quality, then rehearse the defence
- 4Run a dry-run defence with a peer and collect the questions you could not answer
- 5Close those gaps, then defend for real
In the dry run you will be asked what a single request costs, and most people cannot answer. Go instrument it. That question separates a demo from a product.
- Quality on the held-out golden set
- Cost per request and projected monthly cost at realistic volume
- p95 latency under expected load
- Documented failure modes with their containment
Per-target lab variants
Same outcome, five stacks. You train on the one you will actually run.
You must be able to answer
AI use is mandatory on every assignment. The integrity mechanism is defend-your-work, so these are the questions that decide whether the work counts.
- “What is the weakest part of your system, and what would you fix with two more weeks?”
- “What does one request cost, and what would 10,000 a day cost?”
- “Show me where it fails. What happens to the user when it does?”
- “Which decision here would you defend against a senior engineer who disagreed?”
Want to be walked through this phase?
The packs are delivered with live teaching, reviewed assignments, and a defence.
Train with me