Capstone: Ship an Enterprise-Grade AI Product
Brief
Build, harden and operate a multi-tenant SaaS with at least one AI agent feature, then defend it before a panel acting as an enterprise buyer's security, engineering and risk team.
Choose a track
| Track | Product | Hard parts it exercises |
|---|---|---|
| A. Fintech | Instalment/hire-purchase platform: applications, approvals, schedules, payments, reminders | Ledger correctness, idempotent payments, audit, MAS-style resilience, PDPA |
| B. AI Receptionist | Phone/WhatsApp agent that books, answers and escalates for SMBs | Agent security, approvals, real-time streaming, provider failover, cost caps |
| C. Marketplace | Two-sided marketplace with orders, payouts and reviews | Sagas, webhooks, race conditions on inventory, fraud and abuse |
| D. Bring your own | Instructor-approved | Must touch money or regulated data AND an agent |
Minimum requirements
- Multi-tenant with RLS and per-tenant roles; SSO-ready auth
- At least one money or regulated-data flow with idempotency and audit trail
- At least one agent with tool use, approvals for irreversible actions, evals and red-team suite in CI
- Background jobs, webhooks and one saga
- CI/CD with security gates, SBOM, feature flags, sub-1-minute rollback
- OpenTelemetry tracing, 3 SLOs, burn-rate alerts, runbooks
- PITR + immutable off-account backups + restore drill
- Load-tested capacity statement
- Data map, retention, export/deletion, AI register
Deliverables
- Repository with
AGENTS.md, ADRs, OpenAPI spec, runbooks - Architecture and data-flow diagram with trust boundaries
- Threat model (STRIDE + OWASP Agentic mapping)
- Evidence pack for every item in Gates 1 and 2 of Production Readiness Checklist
- Trust package: security overview, sub-processors, AI governance statement, completed security questionnaire
- Live demo environment
The defence (90 minutes)
- Pitch (10 min): product, architecture, risk posture.
- Live red team (30 min): panel attacks the live app: IDOR, injection, prompt injection against the agent, abuse of costly endpoints.
- Incident drill (25 min): panel breaks something unannounced (bad deploy, provider outage, leaked key). Student runs the incident live.
- Procurement interview (25 min): panel questions as CISO, CTO and risk officer.
Rubric
| Area | Weight | Distinction looks like |
|---|---|---|
| Security (app + agent) | 25% | Red team finds nothing high-severity; layered agent defences |
| Correctness and data | 15% | Ledger invariants hold under concurrency; clean migrations |
| Reliability and operations | 20% | Detects drill within 5 min, mitigates within 15, clear comms |
| Delivery | 10% | Small batches, green CI, fast rollback, healthy DORA metrics |
| Governance and evidence | 15% | Every claim backed by evidence; questionnaire credible |
| Architecture judgement | 15% | Knows what was deliberately not built and why (ADRs) |
Automatic fail: secret exposure, cross-tenant data access, or an agent taking an irreversible action without approval during the red team.