FDE case study / decomposition interview

You will get a vague executive request (“make us more efficient with AI”) plus constraints that appear mid-conversation. Interviewers watch whether you decompose the problem, surface risks early, and still propose something shippable.

The C.A.S.E. spine

1. Clarify — success metric, users, data sources, security class, timeline, non-goals
2. Architect — thin end-to-end path from source system to operator workflow
3. Solve the Delta — what the product does not do OOTB; what glue you build first
4. Evaluate — how you prove it works (evals, UAT, adoption), and Day-2 ownership

Palantir popularized decomposition culture for FDSE — see the Palantir interview guide.

Questions to ask in the first five minutes

• What does “done” mean in a number or operator behavior?
• Where is the system of record today (and how dirty is it)?
• Who is the champion — and who can veto?
• What security / residency constraints are hard vs soft?
• What happens if we ship nothing for 30 days (cost of inaction)?

MVP framing that scores well

Propose a vertical slice: one user persona, one workflow, real data (even if incomplete), measurable outcome. Explicitly park gold-plating. Interviewers prefer a boring plan that lands over a beautiful architecture that never ships.

Worked micro-example

“A hospital wants to predict readmission with your platform. Data is on-prem SQL, HIPAA-sensitive, zero cloud today.”

• Week 1: profile data + define “readmission” with clinical sponsor
• Week 2: secure landing zone proposal + DLP / access model (not model tuning yet)
• Weeks 3–4: thin prediction path + eval against historical outcomes + 5-doctor UAT

Notice security and definition-of-done precede fancy ML. That ordering is the signal.

Rubric: junior vs senior answers

• Junior — jumps to code/scripts; thin on stakeholders and security
• Mid — solid MVP; some risk callouts; weak Day-2
• Senior — metrics, blockers, cost, security, and a feedback loop back to product

Practice set

• Bank wants <100ms fraud decisions “using an LLM”
• Manufacturer wants defect reduction with spotty sensor data
• Agency wants an agent over classified docs with intermittent link

Also drill system design and behavioral stories that prove you have lived these tradeoffs. For a deeper discovery and KPI-to-architecture lens, use the Business-to-AI Translation Matrix.