Halvex

AI-powered deal intelligence platform. Halvex is an MCP layer connecting sales and product via Linear, Slack, and HubSpot. It scores live deals, surfaces objection patterns, and closes the loop between what prospects say and what gets built.

What it does

• Deal scores — composite win-probability (text signals + local ML + global Bayesian prior) on every open deal
• Objection engine — extracts objections from notes, maps them to case studies and battle cards
• Product gap loop — links deal friction to Linear issues and notifies sales on deployment via Slack
• Knowledge base — company profile, competitors, case studies, and collateral kept fresh automatically

Local setup

1. Install

npm install

2. Environment variables — copy .env.example to .env.local:

Variable	Required	Description
NEXT_PUBLIC_SUPABASE_URL	✓	Supabase project URL
SUPABASE_SERVICE_ROLE_KEY	✓	Supabase service role key
NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY	✓	Clerk key (pk_test_… dev / pk_live_… prod)
CLERK_SECRET_KEY	✓	Clerk secret key
ANTHROPIC_API_KEY	✓	Anthropic API key
STRIPE_SECRET_KEY	✓	Stripe secret key
STRIPE_WEBHOOK_SECRET	✓	Stripe webhook signing secret
SENTRY_DSN	—	Sentry DSN — omit to disable error tracking

3. Run migrations (Supabase is source of truth for schema):

npx supabase db push        # against remote
# or: npx supabase db reset  # local dev

4. Start dev server

npm run dev

Architecture

src/app/           Next.js App Router pages + API routes
src/components/    React UI (deals, pipeline, dashboard)
src/lib/
  db/              Drizzle schema + typed DB client
  ml/              Composite score, text-signal extraction
  deal-ml.ts       Per-workspace logistic regression
  global-model.ts  Cross-workspace Bayesian prior (nightly cron)
  workspace-brain.ts  RAG knowledge compression
  slack-agent.ts   Slack MCP event handlers
supabase/migrations/  SQL migrations — schema source of truth

Score model

Win-probability blends three components: text signals (always on), local ML logistic regression (activates at 10 closed deals), and a global cross-workspace Bayesian prior (blends down as local data grows, requires 50 pool deals to train). See src/lib/ml/composite-score.ts and src/lib/global-model.ts.

Migration workflow

supabase/migrations/ is the schema source of truth. The Drizzle schema in src/lib/db/schema.ts is kept in sync manually for TypeScript type generation only — do not run drizzle-kit push against production. See CONTRIBUTING.md.

Integrations

Clerk (auth), Supabase + pgvector (DB + embeddings), Anthropic Claude (AI), Linear MCP (product gaps), Slack MCP (closed-loop delivery), HubSpot (CRM sync), Stripe (billing).