Chapter
FDE Delivery
Discovery, scoping, integration, stakeholders, and rollout — the Forward Deployed Engineer interview playbook.
Customer Discovery
FDE interviews start with the customer problem, not the model. Discovery means who hurts, current workaround, constraints (security, IT, data residency), and what "done" looks like in their language.
| Question | Why it matters |
|---|---|
| Who is the economic buyer? | Budget and go/no-go |
| Who is the champion? | Day-to-day access and advocacy |
| What is the workaround today? | Baseline for ROI |
| Where does data live? | Integration feasibility |
| What can we not touch? | Scope and political landmines |
- Shadow a real workflow before proposing architecture
- Collect 10–20 real example queries/docs in week one
- Map systems of record and ACL reality early
- Write problem statement back to the customer for confirmation
Interview signal
Spend the first minutes clarifying stakeholders and constraints. Jumping to RAG architecture without discovery is an FDE fail.
Artifact
Leave discovery with a one-pager: problem, users, success metric, constraints, open questions. It becomes the pilot north star.
Interviews → Workflow map
Tickets/logs → Example queries
IT/security → Constraint list
Champion → Access plan
↓
Problem one-pagerMVP Scoping & MoSCoW
Scoping turns discovery into a time-boxed MVP. MoSCoW (Must/Should/Could/Won't) plus explicit out-of-scope protects the pilot from endless customization.
| Bucket | Example for AI pilot |
|---|---|
| Must | Search 3 approved knowledge bases with citations |
| Should | Slack bot entry point |
| Could | Multi-lingual, agents |
| Won't (this pilot) | Write-back to Salesforce, custom fine-tune |
- Timeline in weeks with demo milestones, not open-ended R&D
- Success metrics and exit criteria written before build
- Single primary persona for v1 — avoid boiling the ocean
- Technical spike on the hardest integration risk first
W1: access + spike connector + golden questions W2: baseline RAG + citations in staging W3: pilot users + feedback loop W4: harden, train, go/no-go review
Scope creep
Every new data source multiplies ACL and quality risk. Park "just one more SharePoint" in Could/Won't until metrics prove value.
Interview signal
Present a crisp MVP and a Won't list. FDEs who cannot say no fail delivery interviews.
Integration & Data Access
Integration is usually the critical path: SSO, connectors, network allowlists, secrets, and ACLs. FDE interviews reward risk-first sequencing — unblock data before polishing the UI.
| Risk | Spike early? | Notes |
|---|---|---|
| IdP / SSO | Yes | Blocks all user testing |
| Doc ACLs | Yes | Wrong ACLs = deal-breaker |
| VPC / private link | Yes | Security review latency |
| Custom UI chrome | No | Defer until data path works |
| Fancy agents | No | After retrieval baseline |
- Inventory systems, auth mechanisms, and rate limits
- Prefer official APIs over brittle scraping
- Design sync: full then incremental; prove deletes work
- Document security review artifacts as you go — do not wait until week 4
Customer IdP ──SSO──► App Customer APIs ─creds─► Connectors → Ingest Network (PrivateLink / IP allow) Secrets manager / KMS Audit logs → customer SIEM (if required)
Interview signal
Name the hardest integration first and how you'd spike it in days, not months. Show respect for customer security process time.
ACL test
Create two test users with different permissions and prove retrieval isolation before any demo to executives.
Stakeholder Management
FDEs navigate economic buyers, champions, end users, IT, security, and legal — each with different success criteria. Misaligned stakeholders kill pilots even when the tech works.
| Stakeholder | Cares about | FDE move |
|---|---|---|
| Economic buyer | ROI, risk | Metric dashboard, business case |
| Champion | Wins & visibility | Quick demos, credit sharing |
| End users | Time saved, trust | Citations, fast UX |
| Security/IT | Control, audit | Diagrams, questionnaires early |
| Legal | Liability, IP | Grounding, retention policies |
- Cadence: weekly working session + biweekly exec update
- Never surprise security with a production cutover
- Translate model metrics into business outcomes
- Escalate blockers with options, not just problems
Champion risk
Single-threaded on one champion is fragile. Build a second relationship in IT or ops before vacation season.
Interview signal
Tell a story where you aligned conflicting stakeholders. Process and communication matter as much as architecture.
Pilot Rollout & Training
Rollout is change management: cohort selection, enablement, feedback loops, and support. A brilliant model with zero training adoption fails the FDE bar.
| Phase | Actions |
|---|---|
| Dogfood | Internal + champion team only |
| Closed pilot | N power users, office hours |
| Expand | Train-the-trainer, playbooks |
| GA handoff | Support model, SLAs, docs |
- Define pilot cohort and success scorecard up front
- Capture failures as eval cases weekly
- Train on when to trust / not trust the system
- Staff a feedback channel with response SLAs during pilot
Users → thumbs / tickets
↓
Triage (bug vs expect)
↓
Fix retrieval / prompt / docs
↓
Re-eval → release notes
↓
Users (trust ↑)Interview signal
Mention training materials and office hours. FDEs who only ship code miss the adoption half of the job.
Demo narrative
Script demos around the customer's real workflow and data — not generic chatbot prompts that hide integration gaps.
Handoff & Productization
Great FDEs leave behind something the customer and core product can own: connectors, configs, runbooks, and patterns — not a snowflake that only you can operate.
| Leave behind | Why |
|---|---|
| Architecture + data-flow diagram | Oncall literacy |
| Runbooks / dashboards | Incident response |
| Eval set + gates | Safe iteration |
| Config vs code split | Customer-specific without forks |
| Training recordings | New user onboarding |
- Upstream reusable connectors to the product instead of one-off scripts
- Document what is temporary glue vs permanent
- Agree support boundaries: vendor vs customer vs SI
- Schedule a formal handoff meeting with acceptance criteria
Hero anti-pattern
Being the only person who can restart the pipeline is not job security — it is delivery failure. Design for your absence.
Interview signal
Describe how a pilot pattern became a productized feature. Palantir/Scale-style FDE interviews love this arc.
Staff depth layer
FDE Delivery — Staff Depth
Forward deployed work is product discovery under enterprise constraints. Staff FDEs sequence risk: data access and ACL truth before model demos, name economic buyer vs champion vs users, define pilot exit criteria in numbers, and leave a supportable system — not a hero laptop deployment.
Anti-patterns → say instead
✕ Demo theater
Happy-path slides win the room then die on SSO, empty indexes, and security review.
Say instead: Week-1 spike on connector + ACL + one golden workflow; demo only what runs on their data.
✕ Boiling the ocean scope
Twelve connectors and custom agents before proving one persona gets value.
Say instead: MoSCoW MVP: one persona, one data source, measurable KPI, explicit out-of-scope list.
✕ No handoff / snowflake forever
Customer can't operate it; every change pages you; deal never productizes.
Say instead: Config over code, runbooks, training, ownership RACI, and a path into the core product.
✕ Ignoring change management
Technically correct pilots fail when users aren't trained or incentives conflict.
Say instead: Pilot cohort, office hours, feedback loop, and exec readout tied to success metrics.
Customer loved a RAG demo on sample PDFs. They want production for 2,000 users over SharePoint with SSO in six weeks.
- Week 0–1: security questionnaire, SSO, site scope, ACL model workshop, success metrics signed.
- Week 2: connector + ACL sync spike; 50 golden questions from champion; baseline retrieval recall.
- Week 3–4: thin UI, citations, feedback, eval CI; load test at expected QPS; threat model review.
- Week 5: pilot 50 users; daily metric review; fix top failure classes only.
- Week 6: go/no-go vs metrics; runbooks; training; support ladder; rollback switch.
- Pilot success example: ≥70% weekly active among cohort, ≥4/5 sats, groundedness ≥85% on golden, <2 Sev-2 incidents, p95 < 5s
- Staffing: 1 FDE + 0.5 SE + customer IT 4h/week; slip if ACL sync blocked
- Kill criteria: no data access by end of week 2 → pause commercial conversation
Timebox discovery risks first; contract success metrics before writing fancy agents.
Cheat sheet
- Economic buyer ≠ champion ≠ end user
- Data/ACL risk first, model second
- MVP + out-of-scope + timeline in weeks
- Signed pilot metrics and exit criteria
- Runbooks, training, support model
- Config > one-off forks
- Rollback and incident path before go-live