Skip to content

05. Assignment 15 — Capstone Project

Week 15. Build the first shippable version of your capstone. Use this with 01_weekly_plan.md, 02_explainer.md, and 03_study_material.md.

Goal

Ship an end-to-end AI system that a senior engineer can review seriously. It does not need final polish yet. It does need clear scope, working integration, first-pass evals, and a credible deployment story.

Required deliverables

1. Product spec

A short document covering: - user persona, - problem statement, - success metrics, - non-goals, - failure risks, - what the demo will show.

2. Architecture pack

  • Component diagram.
  • Data flow diagram.
  • API / interface contracts.
  • One paragraph on failure isolation.
  • Build-vs-buy decisions.

3. Working MVP

At least one happy-path flow must work end-to-end. A stranger should be able to see the value in two minutes.

4. Evaluation pack

  • 10-30 gold queries or scenarios.
  • Pass/fail or scored rubric.
  • One system-level latency budget.
  • One cost tracking sheet or log.
  • One failure-mode list with mitigations.

5. Packaging

  • Reproducible run instructions.
  • Container or equivalent environment definition.
  • README skeleton with architecture, evals, and future work.
  • Rough demo recording or script.

Suggested repository structure

capstone-project/
├── README.md
├── app/
├── prompts/
├── evals/
│   ├── gold_set.jsonl
│   └── run_eval.py
├── docs/
│   ├── architecture.md
│   ├── latency_budget.md
│   └── failure_modes.md
├── Dockerfile
└── Makefile or run.sh

Milestone checklist

  • [ ] Idea locked and scoped.
  • [ ] Architecture diagram written.
  • [ ] MVP works end-to-end.
  • [ ] System eval suite exists.
  • [ ] Latency and cost numbers captured.
  • [ ] README skeleton written.
  • [ ] Rough demo recorded.
  • [ ] External review list prepared for Module 16.

Rubric

Category Weight What strong looks like
Scope judgment 20% Narrow problem, clear success definition
System design 20% Good component boundaries and contracts
Integration quality 20% Happy path works reliably, failures considered
Evaluation discipline 20% System-level metrics, not just vibes
Packaging and communication 20% Demo, README, and diagrams are reviewer-friendly

Non-negotiables

  • Do not hide behind slides without a working path.
  • Do not claim production readiness without instrumentation.
  • Do not use multi-agent complexity unless it earns its keep.
  • Do not skip the eval suite.
  • Do not let scope creep steal demo quality.

Demo script template

  1. State the user and problem in one sentence.
  2. Show the input.
  3. Show system steps quickly.
  4. Show the output and evidence.
  5. Show one metric for quality.
  6. Show one metric for latency or cost.
  7. End with the next production improvement.

External review preparation

Prepare this message now, even if you send it next week:

Hi [Name] — I have built the first version of my AI engineering capstone. It tackles [problem] for [user]. I would value 15 minutes of feedback on the architecture and eval approach. Repo and short demo are here: [link].

Completion gate

Before moving to 06_revision.md, confirm: - [ ] I can explain why I chose this architecture. - [ ] I know the slowest step in the system. - [ ] I know the most expensive step in the system. - [ ] I know the top three failure modes. - [ ] I know what Module 16 must polish next.