Scout

Internal lead-gen operations system for bpulse (work codename: Relay). Runs the Calibrate -> Extract -> Score -> Draft pipeline end to end, with scoring as pure arithmetic (zero model tokens) and a single AI call per lead at draft time.

Pipeline

1. Calibrate (one time per rep) - a quiz plus optional pasted messages build a style card. Every later draft is written in that voice.
2. Extract - raw research becomes structured fields, using the cheapest capable model.
3. Score - a published 12-point rubric in code only. No model is involved.
   • Signal (max 7): hiring, understaffed, funding, stale, weak stack, pain, asking.
   • Completeness (max 5): URL, name, title, specific evidence, verbatim quote.
   • Verdict: send (10-12), research_more (7-9), skip (0-6).
4. Draft - a single strong-model call drafts AND self-checks against two fixed tests (a senior engineer would reply; the message would not survive a company swap). A failed check rewrites the draft once on the same call shape, never ships half-passed.

Guardrails

• No auto-send anywhere. Messages are human copy-pasted; the app logs the text you actually sent so outcomes can be scored.
• Drafts may only claim numbers that exist in the Facts table. Anything else is stripped in code before display.
• No em dashes in any generated text; replaced with hyphens.
• A company marked no or dead is locked for everyone, forever (app-level fuzzy check plus a DB unique index backstop).
• Daily send ceiling per rep (default 15). No follow-up or reply drafting in this pass; the schema and routes are scaffolded, the logic is TODO.

Modes

• demo (default): in-memory store with obvious FAKE seed data and a fixed demo rep. Set GROQ_API_KEY to use the real models, or run without it and the deterministic demo paths cover the whole loop.
• supabase: real Postgres + Auth + RLS. Apply supabase/migrations and set the env vars below.

Models (Groq)

All model calls go through src/lib/ai/routing.ts -> config.ts; no feature code names a model id. By default:

• Extraction runs on openai/gpt-oss-20b (override SCOUT_EXTRACT_MODEL).
• Other structuring tasks (tag, calibrate, classify fallback) run on openai/gpt-oss-20b.
• Drafting runs on openai/gpt-oss-120b and self-checks in the same call.

Override with SCOUT_CHEAP_MODEL / SCOUT_STRONG_MODEL. Free-tier Groq rate limits are handled by a backoff queue that reports "queued, drafting shortly" instead of failing the draft.

Getting started

pnpm install
cp .env.local.example .env.local   # edit as needed
pnpm dev

To move to Supabase mode:

# 1. Point .env.local at the hosted project (URL + anon key).
# 2. Push the schema, RLS, and seed migrations:
supabase link --project-ref <your-project-ref>
supabase db push
# 3. Create the dev auth users via the GoTrue admin API
#    (raw SQL auth inserts can break hosted sign-in):
SUPABASE_SERVICE_ROLE_KEY=<service-role-key> node scripts/seed-dev-users.mjs
# 4. Link reps to those users (idempotent by email):
supabase db push   # applies pending linking migration if not already

Dev users: [email protected], [email protected] (admin), [email protected], [email protected], password scout-dev-password.

Scripts

pnpm dev        # dev server
pnpm build      # production build
pnpm lint       # eslint
pnpm exec tsc --noEmit   # typecheck