Build a production-ready LLM API client with structured output, validation, retries, timeouts, idempotency, tracing, and tests.
Scenario
An internal service must call an LLM to classify support requests. Responses must be structured, repeated requests must be safe, and failures must be measurable and reproducible.
Outcome: Requests and responses have a stable typed shape.
Tasks
Checks
Outcome: The client controls network behavior explicitly.
Tasks
Checks
Outcome: Every call can be explained after execution.
Tasks
Checks
Outcome: Known failure modes can be reproduced locally.
Tasks
Checks
Assessment rubric
Requests, responses, and errors have an explicit typed shape.
Insufficient
The schema is incomplete or errors are swallowed.
Competent
Core contracts are typed and tested.
Strong
Contracts are versioned, and failure modes have dedicated types and tests.
Evidence required
Timeouts, retries, backoff, rate limits, and idempotency are implemented with safe bounds.
Insufficient
Retries are unbounded or retryable and non-retryable errors are mixed.
Competent
Retry and timeout policies are bounded and tested.
Strong
Jitter, a circuit breaker, or multi-model fallback is demonstrated with evidence.
Evidence required
Every call is explainable through request ID, latency, retries, usage, and safe logs.
Insufficient
There is no end-to-end correlation or sensitive data appears in logs.
Competent
Request ID, latency, usage, and redaction are implemented.
Strong
Distributed traces, dashboards, or alert thresholds are implemented.
Evidence required
Success and failure paths can be reproduced automatically.
Insufficient
Only the happy path is tested.
Competent
Unit and integration tests cover the main failure modes.
Strong
Concurrency/load tests and stable regression fixtures are included.
Evidence required