Technical path to become a Forward Deployed Engineer in AI: pin one embedding model, scope before ANN, rail tools/call, eval hit@k in CI, and deploy behind the customer IdP.
Ship
RAG and tools in their VPC, behind their IdP. Same embed dim. Tenant filter before ANN. Reindex on publish.
Fail
Chat wrapper, cosine-only, dim mismatch, uncapped tools, agent loop before hybrid search.
Contract
Structured chunks, pgvector + lexical, rerank 3–5, min-score refuse, allowlist, golden set in CI.
Planes
Runtime (Spring AI / LangChain4j), site (IdP, VPC, PII), control (rails, SLOs, reindex, cost).
Reps
Index a foreign corpus, rail tools/call, break CI on hit@k drop, prove stale-index first.
Done
Explain dim mismatch, show hit@k deltas, deny unknown tools, refuse demos that fail their IdP.