Aangan Enquiry Agent

Turns a phone-call transcript into a qualified, booked, handed-off enquiry for Aangan Studio, and shows the founder a dashboard of what the system generates and what it costs to run.

What it does, in one line: a call comes in, the voice agent asks the right questions, you paste the transcript here, and the app decides Booked / Not the right fit / Needs a person, books a consultation, emails the handoff, adds the lead to HubSpot, and logs it.

What it never does: give a price. A designer shares the figure after seeing the site. pricing.md is never loaded, copied or committed.

The pieces

Piece	Used for
Vaani AI	The voice agent you test with (by hand, no integration)
Gemini	Reads the transcript and writes the handoff note
Cal.com	Books the consultation (test bookings are labelled TEST)
Resend	Emails the handoff, to your own address only
HubSpot	Contact, deal and note for the designers
Neon	The call log (the Dashboard reads from here)
Vercel	Hosts the app

Run a test call, start to finish

1. Open voice-agent-system-prompt.md and paste all of it into Vaani as the agent's system prompt.
2. Make a test call in Vaani. Vaani cannot book or email during the test, and calls cost money per minute, so keep it short.
3. Copy the transcript Vaani gives you.
4. Open the app and go to New call. Paste the transcript, type the call length in minutes, and check the call time (it defaults to now).
5. Press Run this call. You will see each step live: Analysis, Booking, Email, HubSpot, Logged. Each step shows running, done, needs a person, or failed (with what to do next).
6. Check the results.
   • Booked: a Cal.com test booking, an email in your inbox, and a HubSpot deal named TEST - name - area with the handoff note attached.
   • Not the right fit / Needs a person: no booking and no deal. The call appears in Review queue for the front desk the next morning.
7. Review queue: press AI got it wrong if the AI sorted a call badly (you can undo it), and add a callback note. This feeds the AI error rate on the Dashboard.

Where each key goes

Keys live in a file called .env.local in this folder (never committed). .env.example lists the names with blank values. Run npm run check-env to see which are filled, without showing the values.

Name	What it is	Where to get it
GEMINI_API_KEY	Gemini key	Google AI Studio, "Get API key"
GEMINI_MODEL	Model name, e.g. gemini-3.5-flash-lite	Google's models page. Change it here to switch models
RESEND_API_KEY	Resend key	Resend dashboard, API Keys
SELF_EMAIL	The only address emails go to	Your own email (must be the one you signed up to Resend with)
CAL_API_KEY	Cal.com key	Cal.com, Settings, Security, API keys
CAL_EVENT_TYPE_ID	The consultation slot	The number at the end of the event type's address
HUBSPOT_TOKEN	HubSpot Service Key	HubSpot, Settings, Integrations, Service Keys (scopes: contacts and deals, read and write)
DATABASE_URL	Neon connection string	Neon, Connect (use the pooled one)

On Vercel, the same names go in Project, Settings, Environment Variables.

Things you can change without touching code

• prompts/analyse-call.md: the instruction Gemini follows. Edit it like a document: fit rules, how the handoff note is written, the example. Rule: it may never output a price.
• voice-agent-system-prompt.md: what the voice agent says. Keep it under 700 words and free of any price.
• config/app.config.ts: designer names, prices (Gemini, Vaani, dollar-to-rupee rate), the 5-minute target, the average project size used for "pipeline at stake". A price set to null shows as "not set" and is never guessed.

How the numbers are worked out

• Qualified %: qualified calls ÷ all calls.
• Handoffs within 5 min: qualified calls whose handoff email went out within 5 minutes of the call ending (call time plus call length) ÷ qualified calls. For back-dated test transcripts this figure is not meaningful.
• Run cost per booking: all running costs (Gemini tokens, Vaani minutes, free tiers at zero) ÷ bookings.
• AI error rate: calls marked "AI got it wrong" ÷ calls the front desk reviews (not a fit, or needs a person).
• Pipeline at stake: bookings × the average project size, labelled indicative, before conversion.
• Check any figure yourself: node --env-file=.env.local scripts/verify-dashboard.mjs prints the raw counts straight from the database.

Good to know

• Bookings use the first free slot, not the caller's preferred time. The preferred time is passed to the designer in the note and the booking details.
• If Gemini returns something unusable, the app retries once. If that fails too, the call goes to Needs a person and is logged. There are no further retries.
• The handoff note is checked: if it ever contains a price, the answer is rejected.

Adding WhatsApp and the web form later

This version handles incoming phone calls only. WhatsApp and the web form are not built. When you add them, they plug into the same rulebook and the same pipeline:

1. Each new channel only has to turn a conversation into plain transcript text (a WhatsApp thread, or the web-form answers written as "Question: answer" lines).
2. That text goes to the same entry point, POST /api/calls, with a length and a time. Nothing else changes: the same prompts/analyse-call.md decides the outcome, and the same booking, email, HubSpot and log steps run.
3. To tell channels apart on the Dashboard later, add a "channel" column to the calls table (needs a SQL change you approve first).

Run it on your own computer

npm install
npm run check-env
npm run dev

Then open http://localhost:3000.

Safety rules built in

• The app currently has no passcode: anyone with the link can run calls. Add one back before sharing the link widely.
• Emails always go to SELF_EMAIL, from Resend's test sender.
• Everything created in Cal.com and HubSpot is labelled TEST.
• Secrets and pricing.md are in .gitignore and are not in the git history.