Aangan Studio — AI Phone Receptionist

Answers every call, qualifies the lead, books a consultation for good leads, and hands them to a designer with full context. Every call (qualified or not) is logged in the database (Neon Postgres), the dashboard and the CRM.

Live: https://aangan-receptionist-gamma.vercel.app · free voice demo at /talk

Component map

Step	Component	Tool	What it does	Status
Trigger	Customer calls	Vaani (app.vaanivoice.ai) phone agent	Call starts, any hour	✅ Agent created · ⏸ phone number not bought (paid)
Trigger (free)	Browser call	/talk page — Chrome Web Speech API + Gemini	Same agent script, spoken in the browser; free alternative to a phone number	✅ Live
Input	Caller details	Vaani / /talk	Name, area, company, space type, size, budget, timeline, callback number, email, preferred slot	✅
Context	Seed data	seed/services.md, seed/qualified.md	Guide the agent and classifier. pricing.md does not exist and is refused by seed.ts	✅ (placeholder content)
AI	Voice agent	Vaani (Cartesia voice, Gemini brain)	Consent notice first, qualifying questions, never quotes prices or commits	✅
Webhook	Post-call event	POST /api/webhooks/vaani?token=…	Receives call_postprocessing, stores raw event, recovers phone from call_started	✅ Tested
Storage	Call log	Neon Postgres (replaced Supabase)	calls, bookings, crm_mock, webhook_events tables	✅
Processing	Classifier	Gemini (gemini-3.8-flash)	Tags qualified / unqualified + one-line reason + category, extracts fields, writes summary	✅
AI	Booking	Cal.com API v2	Qualified + agreed: books the free slot nearest the caller's preference right after the call (Vaani can't call APIs mid-call)	✅ Real booking tested
Output	Designer handoff	Resend	Email to the round-robin designer: summary, caller details, booking	✅ (test sender → own inbox only)
Output	CRM	HubSpot (service key)	Every lead → contact; qualified → deal (stage: qualified / appointment scheduled)	✅ Tested
Output	Dashboard	Next.js 16 on Vercel	Calls today/week/month, after-hours share, answer rate, lead quality, reasons, leads per designer, tool cost; no phone numbers or transcripts	✅
Code	Repo	GitHub (private)	Push to main → Vercel deploys	✅

Phone call (Vaani)  ─┐                         ┌─ qualified ─→ Cal.com booking → Resend email → HubSpot deal
                     ├─→ save to Neon → Gemini ┤
Browser call (/talk) ─┘                         └─ unqualified → HubSpot contact (logged, never dropped)
                                   ↓
                         Dashboard (Vercel) reads Neon

Cost: ₹0 — every tool runs on a free tier. Only a real phone number would cost money (Vobiz via Vaani: ₹100 setup + ₹500/month, plus ≈₹5.60/min of Vaani call time).

Folder structure

seed/                    services.md, qualified.md (placeholders — the ONLY files the agent/classifier may load)
samples/                 4 sample calls for the "Simulate call" button
db/schema.sql            calls, bookings, crm_mock tables (Neon Postgres)
scripts/                 print-agent-prompt.mts → paste output into Vaani's agent config
src/proxy.ts             basic auth for dashboard + /api/test-call (DASHBOARD_PASSWORD)
src/app/page.tsx         dashboard (no phone numbers or transcripts)
src/app/api/
  webhooks/vaani/        post-call webhook (signature-verified in real mode)
  test-call/             runs a sample through the whole flow synchronously
  tools/availability/    mid-call tool: next Cal.com slots
  tools/book-consultation/ mid-call tool: book a slot
src/lib/
  voice/                 adapter: vaani.ts (real) | mock.ts, chosen by VAANI_API_KEY
  crm/                   adapter: hubspot.ts (real) | mock.ts → crm_mock table, chosen by HUBSPOT_ACCESS_TOKEN
  pipeline.ts            save → classify → book/assign/email → CRM; each step's failure logged, call never dropped
  classify.ts            Gemini Flash, JSON schema output, cost from token usage
  calcom.ts  email.ts  designers.ts  seed.ts  stats.ts  time.ts  env.ts  db.ts

Setup

1. npm install
2. Create a Neon project (neon.tech), open SQL Editor, run db/schema.sql.
3. cp .env.example .env.local and fill in at least DATABASE_URL (Neon connection string) and GEMINI_API_KEY.
4. npm run dev → open http://localhost:3000 → pick a sample → Simulate call.

Anything left empty runs in mock/skip mode and shows as a badge on the dashboard: Vaani → mock webhook, HubSpot → crm_mock table, Cal.com → fake bookings, Resend → logged only. Adding the key to .env (and redeploying on Vercel) switches to the real integration; no code changes are needed.

This machine blocks Next's native compiler (Windows Application Control), so the scripts use --webpack. That works the same on Vercel.

Going live with Vaani Labs

1. npm run agent-prompt → paste into the Vaani agent's system prompt. The first message is the recording-consent notice.
2. Webhook URL: https://<your-app>.vercel.app/api/webhooks/vaani, event call.completed. Put the signing secret in VAANI_WEBHOOK_SECRET.
3. Custom tools (if Vaani supports function calling): POST /api/tools/availability and POST /api/tools/book-consultation (body { start, name, email, phone, call_id }), header x-tool-secret: $VAANI_TOOL_SECRET.
4. Verify the payload: Vaani's docs don't publish call.completed field names, so src/lib/voice/vaani.ts reads several likely keys. Check the first real payload and adjust.

Hard rules: where they're enforced

Rule	Where
Never quote prices	seed.ts allow-lists only services.md/qualified.md and refuses any pric* file; agent prompt forbids it; there is no pricing.md
No commitments / complaints	agent prompt (voice/agent-prompt.ts)
Recording consent	first line of the agent prompt (CONSENT_GREETING)
Unqualified leads never dropped	saveCall runs before anything else; classifier failure → needs_review; CRM sync runs for every status
No keys in code	everything comes from env.ts; .env* is gitignored

Build plan (small steps)

1. ✅ Scaffold, schema, adapters, mocks, dashboard, simulate button
2. Add Neon DATABASE_URL + Gemini key → simulate all 4 samples, check the dashboard + crm_mock
3. Add Cal.com (API key, username, event slug) → check that the qualified sample creates a real booking
4. Add Resend + real designer emails → check the handoff email
5. Deploy to Vercel, set env vars and DASHBOARD_PASSWORD
6. Vaani keys → configure the agent, webhook and tools → place a real test call → fix field mapping in vaani.ts
7. HubSpot private-app token (scopes: crm.objects.contacts.read/write, crm.objects.deals.read/write) → simulate again
8. Polish: call detail page (transcripts behind auth), retry for needs_review, WhatsApp/web-form adapters producing NormalizedCall