Week 23 — Capstone I — Discovery and Architecture

Capstone · Thread A · System design #10

Domain pack: This is where your domain pack pays off: build the capstone on your own industry’s public data.

The capstone

Over four weeks you build one product end to end: discovery → prototype → data → AI → shipped → measured. It should combine at least:

• a user-facing app (Next.js)
• a backend/API with tests
• a data pipeline (Databricks or dbt) feeding analytics
• an AI feature with evals (LLM, RAG, agent or voice)
• production practices: flags, observability, a runbook and an experiment

Pick a theme from your own domain. See the domain packs for public datasets and capstone ideas in pharma and life sciences, healthcare, fintech and e-commerce. Some examples:

• Pharma: a drug-label Q&A and comparison tool over openFDA labels, with a pipeline tracking label changes, and evals for answer accuracy.
• Pharma: an adverse-event signal explorer over FAERS data: medallion pipeline, disproportionality stats, and an LLM that drafts plain-language signal summaries for review.
• Life sciences: a clinical-trial landscape explorer over ClinicalTrials.gov: site and sponsor analytics, plus an agent that answers protocol questions with citations.
• General: extend the course’s ticket service into an AI support co-pilot: triage, RAG answers, escalation analytics.

Use public or synthetic data only. Never use real personal or patient data.

1. Discovery sprint (~3.5 hrs) — swarm plus judgement

1. Frame it using the week 3 format: problem, users, constraints, non-goals.
2. Run an agent swarm. Use five parallel Claude sessions or subagents:
   • competitive landscape
   • user-pain hypotheses from public sources (forums, reviews)
   • regulatory and compliance constraints
   • technical feasibility spike (what’s hard?)
   • a devil’s advocate
3. Talk to at least 3 real people in the target role (or 5 simulated personas, clearly labelled).
4. Synthesise into an opportunity solution tree. Pick one opportunity.
5. Define success:
   • a north-star metric
   • 2 guardrail metrics
   • the eval pass bar for the AI feature
   • the experiment you’ll run in week 26

2. Architecture (~3 hrs)

Write capstone/docs/design.md, the full design doc:

• context and goals / non-goals
• a system diagram showing app, API, database, pipeline, AI, third parties, trust boundaries and sensitive-data flows
• the data model and pipeline (bronze/silver/gold)
• the AI design: model choice, prompts, tools, eval plan, cost and latency budget, fallbacks
• an API sketch
• observability: SLOs, dashboards, alerts
• security and compliance: a STRIDE summary, vendor BAA/DPA needs
• the cost estimate at pilot scale
• risks, and what you’ll cut first

Write ADRs for the 3–5 biggest decisions.

3. Shape the build (~1 hr)

Set an appetite of three weeks at about 8 hours each. Break the work into scopes (Shape Up style), and mark must-haves vs nice-to-haves. Produce:

• a “not doing” list
• an estimation sheet: your estimates now, with the actuals filled in later

4. System design #10 — capstone architecture review (~1 hr)

Present the design doc to Claude as a panel (staff SWE plus data engineer). Iterate on the doc until it survives the review.

5. Check

1. What’s the single riskiest assumption in your capstone, and how will you test it first?
2. Which component would you cut if you lost a week?
3. Where does sensitive data flow in your diagram?
4. What’s the AI feature’s eval pass bar, and why that number?
5. What’s your cost per active user at pilot scale?

6. Teach-back

Write “How I chose my capstone: discovery with a swarm.”

Done when

• Discovery sprint done (swarm, conversations, opportunity tree)
• Success metrics defined
• Design doc and ADRs written
• Build shaped, with a “not doing” list
• Architecture review passed
• Teach-back written