ai-pipeline

name	ai-pipeline
description	AI feature engineering — llama, embeddings, RAG quality, inference reliability, cost/token discipline.
when_to_use	Model wiring, prompt boundaries, retrieval quality, embedding dims, inference failures, routing.

AI pipeline

Role

AI Engineer Lead uses this skill when changing how the product thinks (models, RAG, prompts), not just UI.

Load

• Current stack from [ENGINEERING_REPO]/CLAUDE.md / agents/core/engineering-orchestrator.md (version state)
• Prior long-form playbooks may exist in a bound source org repo; this pack relies on engineering-ai-engineer-lead.md + ai-pipeline instead of _archive/ paths.

Checklist

1. Grounding — RAG answers traceable to corpus; no fabricated OEM specifics
2. Embeddings — dimension and model match Qdrant config and ingest
3. Failure modes — timeouts, empty retrieval, model down — user-visible behavior defined
4. Cost / tokens — reasonable defaults; circuit breakers if applicable
5. Deferred inference migration — llama.cpp → vLLM is roadmap; do not silently change production path without explicit decision

Stop conditions

• External capability claims → security-compliance-evidence + HQ claims matrix

Cursor

Subagents may probe prompts/tests; Mike integrates and validates against evidence.