5dfef33f0e
- intro: cached gemma3 rewrite to one clean paragraph (raw ESCO stays the provenance reference); overlays/<slug>/intro.md wins when present - competences: full essential list, the silent essential[:15] alphabetical cut is gone - hot technologies: derived by p3c from post-gate market ranking (omitted below 30 gated ads), no longer the alphabetical O*NET dump - standardized O*NET proxy caveat for non-exact crosswalk matches - overlays/<slug>/workflow.md preserves curated core workflows across regeneration (flagship overlay included) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PDKeXvpT6tENSvyQGLV1Uq
1.0 KiB
1.0 KiB
- Frame the problem: target metric, baseline, cost of errors, and whether AI is warranted at all (see references/intake.md — mandatory questions).
- Assess and prepare the data: sources, quality, labels, legal basis, leakage risks; build the evaluation set before the model.
- Choose the solution tier: rules/heuristic baseline → prompting an existing model → RAG over private data → fine-tuning → custom training.
- Build the pipeline: versioned data, features/prompts, training or orchestration code, reproducible experiments.
- Evaluate offline against the baseline — overall and per segment; for generative systems run the golden-example eval suite.
- Ship safely: shadow mode → canary → ramp-up, with rollback and guardrails (input validation, output schemas, human fallback).
- Monitor in production: drift, quality proxies, cost, latency; alert thresholds with owners.
- Iterate and govern: retraining triggers, model cards, bias reviews, compliance (e.g. EU AI Act risk class).