Case study · FDE · 40 min
Taking an AI PoC to Production
Plan the path from a flashy demo to a production AI feature with evals, security review, and a 6-week delivery timeline.
FDEdeliveryMVP
Interview opener
Three-week PoC that executives expect to become production.
- 1.IT security timeline is 6 weeks — your move?
- 2.Champion wants flash over foundations — push back how?
- 3.What artifacts remain if you leave day 21?
Strong signals
Hit 0/3
Weak signals (avoid)
- Demo theater without data path
- No success metrics
Requirements
- Context: sales-led PoC chatbot demoed on sample PDFs; customer wants prod in 6 weeks
- Must-have: SSO, ACL, audit logs, eval gate, support runbook
- Success: 200 pilot users; CSAT ≥ 4.0; groundedness ≥ 85%; p95 < 5s
- Constraints: customer VPC; no training on their data; change freeze week 6
Kill scope early
List explicit non-goals (agents, write-back, mobile). Map MoSCoW with the champion in week 0.
Back-of-envelope Estimation
| Metric | Calculation | Result |
|---|---|---|
| Pilot queries/day | 200 users × 5 | 1,000 |
| LLM cost/day | 1K × 3K tokens × $2/1M | ~$6 |
| Eng weeks | 2 FDE + 1 SE × 6 | 18 eng-weeks |
| Security review buffer | 1.5 weeks | parallelize docs early |
| Eval set size | 150 golden Qs | ~2 eng-days to label |
API Design
| Endpoint | Method | Description |
|---|---|---|
| POST /v1/ask | POST | Prod Q&A with SSO JWT |
| GET /v1/health | GET | Dependency checks for customer NOC |
| POST /v1/admin/reindex | POST | Trigger corpus refresh |
| GET /v1/metrics | GET | Prometheus scrape endpoint |
Data Model
| Entity | Key fields | Notes |
|---|---|---|
| pilot_users | user_id, cohort, trained_at | Track enablement |
| eval_cases | case_id, question, expected_docs | Customer-labeled |
| deployments | env, version, config_hash | Staging vs prod |
High-level Design
W0: scope + data access + MoSCoW + threat model W1: SSO + VPC deploy skeleton + sample ingest W2: ACL retrieval + golden eval v1 W3: UX harden + logging/PII redaction W4: security review + load test + runbooks W5: pilot cohort on + feedback loop W6: go-live checklist + hypercare + handoff docs PoC demo ──► staging (customer VPC) ──► prod canary 10% ──► 100%
Deep Dive: PoC Debt to Pay Down
- Replace hard-coded prompts with versioned configs
- Add eval CI before any prompt tweak reaches prod
- Wire real IdP groups — PoCs often fake ACLs
- Document on-call ownership and rollback (previous model/prompt)
Bottlenecks at Scale
- Data access delayed by legal — start DSARs week 0
- Security questionnaire lag — answer in parallel with build
- Champion vacation — identify backup decision maker
Failure Modes & Monitoring
| SLO | Target | Alert |
|---|---|---|
| Pilot CSAT | ≥ 4.0 | Weekly review |
| Groundedness | ≥ 85% | Block expand if < 80% |
| Sev-1 MTTR | < 60 min | Page FDE + customer IT |
| Rollback drill | Passed once | Before go-live |
Related practice topics
Attempt the case aloud first. Then unlock a staff model answer outline.