Pehchaan — never lose a regular
Customer intelligence for Paytm merchants, built from the UPI payments they already receive.
Live demo: https://pehchaan-sigma.vercel.app · Pitch deck (PDF): docs/Pehchaan-Pitch-Deck.pdf · Demo video: docs/Pehchaan-Demo.mp4
- Recognise: every Soundbox payment (payer name + UPI ID) becomes a customer profile.
- Predict: a probabilistic churn model (MBG/NBD + Gamma-Gamma, chosen over BG/NBD and simple rules in an out-of-sample comparison) scores each customer's chance of still being a regular, their expected visits, and their 90-day value.
- Act: a win-back agent picks who's worth an offer, writes to them in Kannada (Gemini), sends it through the Paytm app, and recognises them on the Soundbox when they return (Sarvam voice).
- See the city (Paytm view): when regulars leave many shops in the same week, the network triangulates the cause, e.g. a new quick-commerce dark store.
Paytm Build for India AI Hackathon, Bengaluru, 31 Oct 2026 · Track 1, Merchant Growth AI (with a Track 3 autonomous agent)
Team: Sayantika Laskar, Harsh Shukla
Run
npm install
cp .env.example .env # add free keys (optional)
npm run dev # http://localhost:3000
npm run model:check # fit the models and print the backtest
npm run sarvam:check # one Kannada TTS → STT round trip
npm run validate:cdnow # fit BG/NBD on real CDNOW data and compare with the published parameters
| Key | Get it free | Used for | Without it |
|---|
GEMINI_API_KEY | aistudio.google.com | Writes win-back messages and the morning brief | Kannada templates |
GEMINI_MODEL / GEMINI_THINKING | optional | Default gemini-3.5-flash-lite (falls back to gemini-3.6-flash on quota), thinking low | — |
SARVAM_API_KEY | dashboard.sarvam.ai | Soundbox voice in Kannada / Hindi / English (Bulbul v3) and spoken questions (Saaras v4) | Browser voice; tap-to-ask questions |
MONGODB_URI (+ optional MONGODB_DB, default pehchaan) | MongoDB Atlas free tier | Live phone loop: offer links, opens, redemptions and opt-outs shared across servers (offers expire after 7 days) | Upstash if set, else in-memory (works locally only) |
UPSTASH_REDIS_REST_URL / _TOKEN (or KV_REST_API_URL / _TOKEN) | Vercel → Storage → Upstash Redis | Alternative store for the live loop | In-memory store |
Every Gemini call is cached and raced against a timeout, falling back to a template, so the demo never stalls on the free tier.
Proof, not vibes (npm run model:check)
All numbers come from a deterministic simulation of one HSR Layout kirana: 420 customers, 14k payments, Nov 2025 → Oct 2026, with a dark store opening on 18 Aug.
| Check | Result |
|---|
| Backtest: train to 1 Jun, predict 1 Jun–15 Aug | 2,882 predicted vs 2,899 actual visits (−0.6%) |
Who-left ranking (AUC, npx tsx scripts/compare-models.ts) | 0.95 on the shop · 0.77 on real CDNOW data (plain BG/NBD: 0.32 on CDNOW) |
| Per-customer correlation | 0.92 |
| Top 10% of customers by model | made 41% of holdout visits |
| Visits vs expectation after the opening | −47% within 500 m · −19% at 0.5–1 km · ~0% beyond 1 km |
| Network radar | 7 of 36 shops flagged; change point in the week of 17 Aug; source located within ~360 m |
Real data — CDNOW (npm run validate:cdnow) | Fitted parameters within 3% of Fader, Hardie & Lee (2005): r 0.242 vs 0.243, α 4.394 vs 4.414, a 0.814 vs 0.793, b 2.477 vs 2.426. Holdout: 1,631 predicted vs 1,882 actual repeat purchases (−13%) |
The LLM never decides who or how much. It only writes the words. The win-back agent picks the offer size by Thompson sampling over ₹20 / ₹30 / ₹50 and learns from every redemption.
Screens
| Route | What it shows |
|---|
/ | Landing: 3D product preview, pinned scroll story, proof, business model, FAQ |
/app | Today: KPIs, visits vs model expectation, Kannada morning brief, slipping-away list, counter recognition |
/app/customers | Every customer, with segment, probability of still being a regular, and value; drawer with the decay curve and household |
/app/winback | The agent: plan → Kannada messages → send → simulated returns and ₹ recovered |
/app/model | Backtest chart, fitted parameters, a worked example, and where the LLM is (and isn't) used |
/app/radar | Paytm view: map with a week-by-week scrubber, inferred dark store, and a ground-truth reveal |
/app/import | Your data: upload a Paytm payment export (CSV), map columns, score your own customers — in the browser |
/app/trust | Trust centre: RBI / NPCI / DPDP / TRAI principles, what the LLM sees, consent ledger, audit log |
/app/business | Business case: interactive calculator of value to Paytm and to one merchant |
/kiosk | Full-screen Soundbox: live payments greeted aloud, voice questions, Kannada / Hindi / English |
/o/[code] | The customer's phone: Paytm-style offer with Redeem and Stop offers |
Edge cases: lockdowns, outages, no network, can't visit
| Situation | What Pehchaan does | Proof |
|---|
| City-wide shock (COVID-style lockdown, floods, bandh) | Paytm's network sees ≥ 50% of shops drop the same week with no spatial source → classified city-wide. The model switches to business time (lockdown days count only for the trade that happened), discount offers pause, the agent switches to WhatsApp ordering + delivery. | 34/35 shops, spatial R² 0.00 → city-wide. Worst-week false churn alarms 47 → 3 (npm run resilience:check) |
| Local shock (a dark store opens) | Few shops, strong distance pattern → local: real churn, win-back offers and the radar. | 6/35 shops, R² 0.71 → local |
| Soundbox / network outage | One shop at zero while peers trade → outage: its scores freeze, events backfill with real times when it reconnects. | False alarms 18 → 1 |
| No network at the shop | Offline-first PWA (service worker), churn model runs in the browser, actions queue in an outbox and sync on reconnect; offers are minted on-device and carried inside the QR. | DevTools → Offline → Run agent |
| Rural, zero bars | Ad-hoc local links: offer in the QR (shop → customer), redeem by sound (customer → shop, 2-second chirp, open-source ggwave, MIT), store-and-forward sync; offline UPI rails (UPI Lite X, UPI 123PAY) count as visits when they settle. | Tested: chirp decoded in-browser → win moment |
| Shop with no tower at all | Pehchaan Relay (/app/relay on the shop, /relay on any phone): offline actions are sealed into one bundle (ECDH P-256 + AES-256-GCM to Paytm's key, ECDSA-signed by the shop device), handed to any visiting phone as an animated QR, and delivered when that phone finds signal. Paytm verifies, de-duplicates by action id, applies each action once with its original time, and returns a signed receipt so the shop can forget it. In production the same bundle also travels over Bluetooth / Wi-Fi Direct (Nearby Connections) inside the Paytm app. | npm run relay:check: delivered 4 h late, receipt verified, second carrier → duplicate, tampered bundle → rejected; browser rehearsal: offline Soundbox redemption → phone scan → delivered → shop clears its queue on reconnect |
Live QR loop (show it to mentors)
- Open
/app/winback on the laptop and press Run agent. A QR appears in the Paytm notification card.
- Scan it with any phone: the offer page opens and the dashboard shows Opened on a phone within ~2 s.
- Tap Redeem at the store: the laptop plays the win moment and the Soundbox greets the customer by name. Stop offers records a consent withdrawal.
On Vercel this needs MONGODB_URI set so every serverless instance shares the same events.
3-minute demo
/ (30s): "Quick commerce is taking the kirana's best customers, and the shopkeeper can't see who's leaving."/app (40s): visits fall below the model's expectation from mid-August. Press Play on Soundbox to hear the morning brief in Kannada./app/customers (30s): open a slipping customer; their probability of still being a regular decays from 80% to 31%./app/winback (40s): Run agent, then Hear it, then Simulate the next 7 days to show who came back and the ₹ recovered. Back on Today, Simulate payment: "Aisha is back after 48 days."/app/radar (40s, the wow moment): press play; the dots turn red week by week, the inferred store appears, then Reveal ground truth: about 360 m off.
Open /app and /app/winback about 30 seconds before presenting so Gemini's answers are cached.
Code map
src/lib/pehchaan/sim.ts: deterministic customer and payment simulation, including the dark-store shocksrc/lib/pehchaan/btyd.ts: BG/NBD and Gamma-Gamma (maximum-likelihood fit with Nelder–Mead), P(alive), expected visitssrc/lib/pehchaan/model.ts: scoring, segments, win-back priority, backtest, shock analysissrc/lib/pehchaan/network.ts: multi-merchant baselines, CUSUM change points, spatial triangulationsrc/lib/ai.ts: Gemini (Interactions API, JSON schema) + Sarvam TTS, with cache, timeout and templates
The data is simulated for the demo; the models are real and run unchanged on real Paytm payment histories.
