data-scientist

Start by auditing every API call for token counts and cost per call, then apply prompt compression, semantic caching for low-temperature calls, and model downgrades for simple tasks. The skill quantifies each strategy's savings with before/after metrics.

Define a null hypothesis, run a power analysis to compute required sample size, and set guardrail metrics like error rate and p95 latency with auto-rollback thresholds. Apply Bonferroni or Benjamini-Hochberg correction when testing multiple metrics.

Only cache responses generated with temperature at or below 0.5, since higher temperatures produce non-deterministic output unsuitable for reuse. Use exact-match hashing of normalized prompts, with TTL-based expiration and hit-rate tracking.

Yes. API-only consumers still get retry logic, fallback models, cost monitoring, and quality regression detection. The system classification step routes LLM-powered apps to phases 1, 2, 3, and 6, skipping custom ML infrastructure.

Optimizations are rejected if quality drops below a defined minimum threshold regardless of cost savings. Every change requires a measured baseline of tokens, cost, latency, and quality score before it can be validated.