AI EngineeringInterview preparation
Intermediate·Concept·2 min read

Project 9 — Deterministic plus agentic workflow

Practice this question

Requirements

Interpret input, prepare a draft, validate required fields, wait for review, execute once, and notify reliably.

Architecture

AI interpretation → deterministic state machine → durable approval wait → idempotent service → outbox notification.

Technologies

Existing AI/graph layer plus Temporal or another justified durable workflow, or BullMQ with explicit application state for simpler jobs.

Database design

Draft versions/hashes, approval records, operation keys/receipts, workflow status, and outbox events.

API design

Draft create/edit/review/submit/status; require expected version and operation key; expose uncertain outcome explicitly.

AI design

Language extraction and draft wording only; code controls calculations, permissions, retries, and transitions.

Security

Authenticated exact-action approval, current-rule checks at commit, least-privilege activities, safe recipient validation.

Evaluation

Duplicate messages, worker crashes, stale approvals, transaction/dispatch gaps, external timeout, and compensation limits.

Observability

Workflow events, wait age, retry counts, receipt links, unknown outcomes, and notification lag.

Deployment

Durable storage and timers, worker recovery, schema/version migration, redrive/reconciliation operations, and backup restore test.

Interview questions

  • “Exactly once?” Answer: design effectively-once business effects through deduplication and reconciliation; do not assume the network guarantees it.
  • “Can every action be undone?” Answer: disclosure and email often cannot.

Acceptance: The same approved logical request creates one effect across retries and restarts.

Adapted from 13_AI.MD · “Project 9 — Deterministic plus agentic workflow”. Code examples are study fragments. Difficulty and core labels are editorial estimates; version-specific behavior should be checked against your project.