Illustrative sample — not a real person

Senior Backend Engineer — growth-stage SaaS

Passport № SAMPLE-0001 Evaluated July 2026 Refresh recommended after 12 months Identity verified (document + employment claims, level 2) Overall Strong evidence

Work context observed

90-minute live session: design, implement and debug a subscription-event processing system (webhook ingestion, idempotency, retry policy, billing-state reconciliation) against a seeded codebase with two planted defects. AI tools permitted and observed throughout.

Evidence by dimension

Problem framing

Strong evidence

Asked about delivery guarantees and duplicate-event volume before writing anything; explicitly narrowed scope to at-least-once delivery with idempotent consumers.

"Before I pick a dedup strategy — do we control the webhook sender's retry policy, or is it a third party? That changes where the idempotency key has to live."

Architecture & trade-offs

Strong evidence

Proposed queue-backed ingestion with an outbox pattern; articulated the cost of exactly-once claims and chose verifiable idempotency instead. Compared Redis- vs DB-backed dedup with concrete failure scenarios.

Debugging under pressure

Strong evidence

Found both planted defects (a race on subscription-state update; a swallowed exception in the retry path) in 31 minutes, narrating hypotheses and eliminating them with logs rather than guesswork.

Use of AI tooling — judgment & verification

Strong evidence

Used an assistant to draft the reconciliation query, then caught its incorrect join condition by testing against the seeded edge case before accepting it.

"The generated query looks plausible but it double-counts grace-period rows — let me check against the cancelled-then-renewed case first."

Edge-case & correctness discipline

Moderate evidence

Handled clock-skew and out-of-order events when prompted, but did not raise them unprompted. Evaluator note: consistent once surfaced; recommend probing in the team interview.

Communication of decisions

Strong evidence

Kept a running decision log unprompted; final walkthrough was ordered, complete, and honest about the shortcut taken on metrics emission.

Evaluator provenance

  • 9 years backend (payments & subscription systems)
  • Calibration: 96% agreement on benchmark set (n=14)
  • Conflicts declared: none (no referral relationship)
  • This session second-scored: yes — agreement on 5/6 dimensions

Artifacts attached

  • Session recording (candidate-consented)
  • Final code diff + decision log
  • Structured scorecard (rubric v2.3)
  • Second-scorer notes

Candidate response

Candidate comment (unedited): "Fair report. On edge cases: I'd flag that the session brief said to prioritize the happy path first — I'd normally spec out-of-order handling in the design doc before implementation."

Sharing & consent

  • Owned by the candidate; shared per-employer with explicit approval
  • Current-employer category blocked from discovery by candidate setting
  • No personality, emotion or protected-characteristic inference — by policy, ever
  • Correction & appeal: factual-error challenge with human review
  • Retention: deleted on candidate request; auto-review at 24 months