MS Enterprises ERP
AI-native ERP for a precision die & machining job-shop. Multi-tenant (Postgres RLS), TypeScript end-to-end.
Design docs: see docs/ (00–06) and the published blueprint. This is the Phase-1 implementation.
Stack
Next.js 15 · TypeScript · PostgreSQL + Row-Level Security · Drizzle ORM · MinIO · Redis · Claude (AI) · Docker Compose.
Quick start
pnpm install # install workspace deps
pnpm infra:up # start Postgres + Redis + MinIO (Docker)
pnpm db:setup # create app role + RLS, run migrations, seed a demo tenant
pnpm dev # run the web app → http://localhost:3000
Demo login (seeded): [email protected] / password123
AI features (optional — set ANTHROPIC_API_KEY in .env)
- ⌘K assistant (ask & act) — ask your data in English or हिन्दी ("top 5 customers by sales", "इस महीने कितनी बिक्री हुई?"), and operate the whole ERP by chat: record and update leads, log follow-ups, convert leads, manage customers, draft quotations with priced line items, move quotation/invoice statuses, convert quotations to invoices, raise invoices, and open any screen or print-ready PDF. Every write is staged as a confirmation card (validated + permission-checked server-side, stored in
ai_action as an audit trail) and runs only after you click Confirm — GST math and document numbering stay deterministic in code, never from the model.
- Smart quotation drafting — describe the job; the AI proposes line items, rates (anchored on your quote history), terms, assumptions and anomaly flags. You review and edit everything; totals & GST stay deterministic in code.
- Terms polish — one-click drafting/tightening of terms & notes on quotations and invoices.
- Pricing memory — while typing an item on a quotation/bill/order, past lines pop up with the last rate, HSN and unit (one tap fills the row); the assistant's
price_history tool anchors new quotes on real history ("last time ₹30,000 for Sharma Auto").
- Messages you send, drafted for you — "send a payment reminder / follow-up / delivery update on WhatsApp": the assistant writes it, the card shows WhatsApp / email / copy buttons, and a public share link to the quotation or bill PDF (signed, expires in 60 days, no login needed) is inserted. The app never sends anything itself.
- Today's briefing — the home screen carries a 3–5 line plain-words summary (money overdue, follow-ups, quotations waiting, orders due, how the month is going), written once a day per business from the same numbers the dashboard shows.
- Email enquiries read by AI — free-form enquiry emails that the IndiaMART/TradeIndia parsers don't recognise are read by the model into a draft (name, phone, requirement) for the review screen; newsletters and notifications are parked, recoverably.
- AI activity & usage (
/settings/ai, owner) — every proposal the AI made and whether it was saved, tokens per day, by feature and by person.
Model routing, cost metering (ai_usage table), prompt caching and guardrails are described in docs/05-ai-integration.md. Without a key, AI surfaces show a setup hint and everything else works normally.
Also: global search (top bar or /), customer statement-of-account and payment-receipt PDFs, dark mode (Auto/Light/Dark in the sidebar).
Layout
apps/web Next.js app (UI + server actions + AI routes)
packages/ai Claude gateway: model routing, SQL guard, structured drafting
packages/db Drizzle schema, migrations, RLS, tenant client, seed
packages/core shared types, zod, enums, utils
docs/ product & technical blueprint (00–06)
Database: local Docker Postgres by default; swap DATABASE_URL for managed Postgres (Neon / Supabase free tier) with zero code changes — see DEPLOY.md.
Multi-tenancy
Every business row carries tenant_id; Postgres RLS enforces isolation. The app connects as a non-superuser (app_user) and every request runs inside a transaction that sets app.current_tenant — a query physically cannot read another tenant's data.
