Build-session scaffold
Auth, database, model access, and deploy are already wired. Pick the archetype that matches the scenario, copy its kickoff prompt, and start building the AI layer first.
System status
checking…
Kickoff prompts
RAG chatbot
Synthetic corpus → hybrid retrieval → grounded, cited answers with refusal. Mini Recall@5 eval included.
Agent with tools
Typed read-only tools over seeded Postgres data, visible tool trace, one gated write, deterministic tool evals.
Multi-agent flow
Analyst → writer → critic with bounded revision loop. Code owns control flow; models own judgment.
Structured extraction
Messy synthetic docs → schema-constrained extraction with evidence quotes, needs-review lane, field-level eval.
Already working end to end
/chat — signed-in user → streaming Claude call → run logged to
perficient.runs in Neon. The full pipe (auth → API → model → database) is proven before the clock starts.