Aangan Enquiry Desk: AI phone-enquiry system for Aangan Studio, Pune
MESA Case 03. An AI agent answers every incoming enquiry call within seconds, day or night, qualifies the caller against Nikhil's rubric, never quotes a price, hands qualified leads to a designer with full context, books a free consultation, syncs to HubSpot and shows the studio an executive dashboard.
How a call flows
Caller -> Vaani voice agent (answers in seconds, 24x7)
-> webhook to this app (FastAPI on Vercel)
-> Gemini judges the call: 5-criteria verdict, gist, mood, key moments
-> Neon Postgres stores the call, transcript, audit trail
-> qualified: designer chosen by workload and area
-> designer email (gist, answers, "don't re-ask" list, one-click Accept / Listen)
-> Cal.com books the nearest free slot, client gets the calendar invite
-> every caller: HubSpot contact + note; qualified: HubSpot deal
-> dashboard updates (KPIs, kanban, escalations, forecast)
Requirements from the case and where they live
| Requirement | Implementation |
|---|
| Answer within 5 minutes, day and night | Vaani agent picks up in seconds; dashboard shows median/slowest answer, % under 5 min, after-hours calls |
| Qualify per Nikhil's rubric | knowledge/qualified.md + knowledge/services.md drive the prompt; Gemini returns a structured 5-criteria verdict (app/prompts.py) |
| Never quote prices | The price table is never given to the model; app/guard.py blocks any price figure before it is spoken; the deflection line steers to the free consultation; dashboard counts deflections |
| Handoff to designers with context | app/integrations.py designer_email: gist, who/what/where/timeline, price question status, "don't re-ask" list, transcript, signed Accept and Listen links |
| Book a consultation + invite to client | Cal.com v2 (app/slots.py): booking made post-call from the caller's preferred time; Cal.com emails the invite. .ics fallback |
| HubSpot sync | Contact for every caller, deal only for qualified leads (amount is an internal estimate), call note with dashboard link |
| Executive dashboard | app/static/dashboard.html: AI digest, 4 KPI cards, kanban (drag or arrows), escalation feed, pricing-deflection tracker, response speed, designer workload, mood heatmap, off-hours revenue ticker, forecast, win chance, first-touch timers, NL search, call drawer with transcript, waveform markers and audit log, CSV/PDF export, system health |
| Running cost and ROI | Per-call cost (Gemini tokens, telephony, speech) and hours/money saved on the dashboard; rates and project-value assumptions are editable settings |
Accuracy check
replay.py runs the 20 real September 2026 call transcripts (kept out of this repo because they contain customer names; put the file at data/enquiries_sep2026.txt to run it) through the same pipeline. Result against my reading of the rubric: 20 of 20 verdicts match (judgement calls are listed in replay.py).
Stack
FastAPI (Python), static HTML/JS dashboard, Gemini (gemini-3.5-flash), Vaani (voice), Neon Postgres (SQLite locally), Cal.com, HubSpot, Vercel (+ weekly cron), GitHub. Emails via Resend.
Run locally
cp .env.example .env # add GEMINI_API_KEY; LIVE_INTEGRATIONS=0 keeps HubSpot/email/Cal.com off
./run.sh # dashboard http://localhost:8000, call simulator /simulator
.venv/bin/python replay.py
MOCK_LLM=1 runs the whole pipeline without any API key (canned replies, for plumbing tests).
Honest status and decisions
- Verified end to end with real Vaani voice calls (browser web calls). One call produced: transcript and Gemini verdict, auto-assigned designer, HubSpot contact and deal, a Cal.com booking (status accepted) from the caller's preferred time, a warm client email and a designer email (both via Resend), and a dashboard update. An inbound phone number is not attached to the Vaani agent yet, so the calls so far came from Vaani's web test, not the phone network.
- Email runs through Resend (
RESEND_API_KEY, Gmail SMTP still supported). Delivery verified. Without a verified sending domain Resend only delivers to the account owner's address, so all designer and escalation emails currently go to STUDIO_HEAD_EMAIL (the owner). To email real designers and clients, verify a domain at resend.com/domains and set RESEND_FROM.
- Designers are placeholders in
designers.json (name, email, optional areas, commercial). Placeholder addresses fall back to the studio head.
- Pricing rule wins over the brief. The brief mentions per-sq-ft quotes, but
pricing.md forbids any number, so the agent never quotes one.
- Open question for Nikhil:
services.md caps commercial work near 3,000 sq ft with no minimum; a 500 sq ft minimum appears only in transcript T18.
- Mood is read from the transcript text, not vocal tone. Win chance is a transparent rules-based estimate, not a trained model. Cost figures are estimates from configured rates, not vendor bills.
LIVE_INTEGRATIONS=0 locally is a safety switch so tests never touch the real HubSpot, Cal.com or inbox.
