Customer discovery to production MVP wedge

Expected question

"A Fortune 500 wants to deploy our AI platform in their environment. Walk me from discovery to a production MVP — not a demo."

Variant forms

• "Design a private VPC RAG for a healthcare customer with HIPAA and 50M documents."
• "Customer has Okta SSO, Snowflake data, and wants agents in their AWS account. Deploy it."
• "How do you go from POC to production without a permanent custom fork?"
• "What's your week-1 discovery checklist?"
• "How do you score which use case is the first wedge?"
• "Customer data is split across SAP, Salesforce, and Postgres. How do you unify for an agent?"
• "When do you fine-tune vs RAG vs prompts for this customer?"
• "How do you measure MVP success with the customer?"
• "How do you stage a rollout so the customer sees value in week one, not month six?"
• "Tell me about shipping in days something that 'should' have taken months — which corners did you cut?"
• "How do you demonstrate ROI with before/after metrics a sponsor will trust?"

Where this actually gets asked

OpenAI FDE, Databricks FDE, Cohere enterprise deploy, and Anthropic Applied AI “get to production” rounds. Interviewers want a scored wedge under constraints, not a reference-architecture dump.

The question, as it might actually be asked

"Don't give me a reference architecture. Give me the first shippable wedge and how you earn the next one."

The framework

30-second thesis

I'd start by discovering the job and the constraints I can't change, score a few candidate use cases, and ship a read-first walking skeleton under their SSO with HITL on irreversible writes. Rollout expands only when evals say so. I hand off ownership — and I won't leave a permanent fork of core product in their account.

2-minute method

Discover → score → wedge → integrate → govern → eval → handoff. Spoken version:

Week one I'm not picking models. I'm asking: who waits on what decision, what's system of record, how identity and ACLs actually work, what's irreversible day one, and what one number with a baseline means “this worked” in 90 days. Then I score use cases — value × feasibility × risk × reuse × whether a sponsor will show up weekly — and I pick one.

Default wedge: read/assist first, access before ranking, gateway + HITL for side effects, eval gate before broad rollout. Config and connectors outside core. No permanent customer fork.

That's the same brand I ship publicly — governed agents, access-aware RAG, HITL, evals — wired into their VPC or private link, not a parallel demo stack.

Week-1 discovery checklist

1. Jobs to be done: who waits on what decision today?
2. Systems of record: which write is authoritative?
3. Identity: SSO, groups, row/object ACLs for retrieval.
4. Data reality: freshness, PII, residency, quality of labels.
5. Irreversibility: which actions need HITL day one?
6. Change control: freeze windows, CAB, who can approve prod?
7. Success metric: one number with a baseline window.
8. Exit: who owns the system when FDE leaves the critical path?

MVP scoring rubric (pick highest)

Wedge design defaults

• Read path first (RAG/assist) before autonomous writes.
• Access-before-ranking for enterprise retrieval.
• Gateway + HITL for side effects (don't invent a second policy system per customer).
• Eval gate before broad rollout: groundedness, tool success, HITL reject rate.
• No permanent customer fork of core product — config, connectors, policies.

Requirements

Functional

• Named sponsor, workflow, and 90-day metric.
• SSO-authenticated users; ACL-aware retrieval where enterprise docs are involved.
• Explicit write boundary (HITL or blocked) for irreversible tools.
• Staging → shadow → canary path.

Non-functional

• Residency / egress story; audit to customer SIEM where required.
• Cost/latency budgets per successful task.
• Handoff criteria written before go-live theater.

Core entities

• Wedge: thinnest production path that moves the metric.
• Connector: customer-specific adapter; must not fork core.
• Policy / gateway: authorize side effects; meter; audit.
• Eval pack: golden set + online monitors owned after FDE exit.
• System of record: destination of any approved write.

Data / process flow

Rendering architecture diagram…

1. Authenticate via customer SSO.
2. Retrieve under ACL predicates (access before rank).
3. Assist / cite; decline when unsupported.
4. Side effects only through gateway + HITL.
5. Emit audit; promote only when eval gates pass.

High-level design

Users → Customer SSO → App / agent runtime (customer VPC or private link)
                      → AI gateway (policy, HITL, quotas)
                      → Retrieval (Snowflake/docs under ACL)
                      → Model (private endpoint | allowlisted egress | on-prem)
Audit → customer SIEM

Deep dive 1: fine-tune vs RAG vs prompt (R)

Deep dive 2: unify SAP + Salesforce + Postgres without MDM theater

I'd pick one workflow's entities, read APIs/CDC for that slice, and build a canonical projection for agent tools only. Writes go back to the system of record with idempotency + HITL. I would refuse to “unify the enterprise” in wave one.

Deep dive 3: POC → production without a permanent fork

Config, connectors, and policies live outside core. Dual-run / shadow before cutover. Kill switch and previous prompt/index bundle warm. Anything that required a core fork is product debt with an expiry — not a badge of heroism.

Quantitative trade-offs

Migration and rollout

1. Week one: SSO + one data source + read-only assist in staging.
2. Shadow beside humans; collect golden failures.
3. Canary cohort (illustrative H: 1% → 5% → 25%) with holdback — label as assumption unless measured.
4. Enable HITL writes only after tool-success and reject-rate gates.
5. Handoff runbooks; FDE exits critical path when named owners pass a drill.

Org ownership

• Sponsor owns metric and weekly review.
• Security / identity owns SSO, secrets, egress exceptions (with expiry).
• Platform owns gateway invariants and connector SDK.
• Domain ops owns HITL queue and prompt/eval content.
• FDE owns critical path until exit criteria met.

Situation

At Lucid (P), Supply Chain / Commerce partners need production AI on exception and ops workflows — not a lab demo. Parallel interview framing is Fortune-500 deploy of “our platform” into their identity and data plane.

Task

Get from discovery to a production MVP wedge: scored use case, integrated into real systems, governed irreversible actions, measurable adoption, handoff — without boiling the ocean or forking core product permanently.

Action

1. Run the week-1 checklist; refuse to pick tools before jobs, systems of record, and metric.
2. Score candidates; pick highest value × feasibility × sponsor availability.
3. Ship read-first assist with access-aware retrieval; HITL on irreversible paths.
4. Integrate via existing event/identity plane rather than a parallel stack.
5. Gate expansion on stratified evals and HITL reject trends.
6. Document what is connector/config vs what must become product; leave owners named.

Method proof for panels (O): Enterprise RAG, AegisAI, VAP, Content Factory — same brand under customer VPC/account boundaries. Not a claim Lucid runs those repos.

Open reference wedge — Acme Support Agent Embed (O)

When a panel asks “show me a 90-day embed under constraints,” I point at the named wedge on the existing spine — not a thirteenth SaaS:

Spoken: venkat-ai.com/fde Embed lab · ADR-032 · operator wiring. Acme is a reference customer name, not a logo claim.

Result

A deploy narrative that survives Staff+/Principal panels: wedge under constraints, eval gate, handoff. Attributable savings or hours only with baseline windows (P when you have them; otherwise H).

The follow-up question you should expect

"What did you cut to ship in days instead of months?"
Multi-agent autonomy, multi-source MDM, and fine-tuning. Kept SSO, one source, citations, HITL on writes, and a golden set. Corners cut on polish and breadth — never on authorization or irreversible-action governance.

What I'd ask them

1. Who owns the 90-day metric and shows up weekly?
2. What's irreversible on day one, and is there HITL capacity?
3. Can we get staging SSO + one data slice this week — or is that a six-month ticket?
4. What dies if we fork core for this account?

Candidate-owned evidence prompts

1. Which Lucid wedge will you name (exception type / Commerce path)?
2. What is your 90-day metric and baseline window?
3. Which three non-goals will you say aloud?
4. Who is the handoff owner for prompts/evals after you leave?

Author reference (do not memorize)

STAR is a template. Upgrade O→P only with evidence-ledger facts. Hosted vs self-hosted detail lives in 07.

Staff+/Principal signal rubric

• Mid-level: Lists tools (LangChain, vector DB) as the plan.
• Senior: Names SSO, data source, and a v1 use case.
• Staff+: Scoring rubric, non-goals, HITL boundary, eval gate, handoff owner.
• Principal: Productization path (platform vs customer-specific), FinOps, multi-wave adoption.

Related

• 02-ambiguous-customer-decomposition.md
• 07-enterprise-vpc-sso-hosted-vs-selfhosted.md
• ADR-032 Acme Support Agent Embed
• ../ai-system-design/02-rag-platform-at-scale.md
• ../cloud-architecture/09-enterprise-ai-landing-zone.md
• ../ai-system-design/07-llm-evaluation-observability-platform.md