Pujo Pilot

Durga Puja intelligence layer
PROOF OF CONCEPT · SIMULATED DATA

Which pandal, right now, without the two-hour line

Rows marked LIVE carry a queue length reported in the last 45 minutes, either by a committee volunteer at the gate or by geofenced visitors, in minutes rather than a normalised score. GOOGLE rows take their hourly shape from Google Maps Popular Times. MODELLED rows have no coverage and fall back to a curve. Festival uplift across Panchami to Dashami is still modelled, not measured.

Modelled for
8:00 PM

Why this is a business, not a hackathon project

₹50,000 cr
Bengal's Pujo economy

British Council's 2019 study put the creative economy of Durga Puja at roughly ₹32,000 crore, ~2.6% of the state's GDP; trade bodies now estimate well above ₹50,000 crore across five days. UNESCO listed it in 2021. The footfall exists. The software does not.

3,000+
Registered pujas in Kolkata alone

Roughly 250 are big-budget with paying sponsors, ticketed hospitality and police coordination obligations. Those are the ones that already buy LED walls, drones and security agencies — and have nothing telling them where the crowd is.

3 buyers
One data asset

Committees pay for ops and sponsor proof. Kolkata Police and the tourism department pay for the city-wide view. Brands pay for verified footfall attribution. The consumer app is the sensor network that makes all three sellable.


What is real in this demo

The live queue feed is a working system, not a mock: committee desks and geofenced visitors report a wait in minutes to a Cloudflare Worker, reports decay out after 45 minutes, and a weighted median stops one bad report moving the number. Where nobody has reported, approach-road congestion fills in and is badged as a proxy. Five pandals also carry a measured hourly curve pulled from their own Google Maps listing: Santosh Mitra Square, Sreebhumi, Ekdalia Evergreen, Deshapriya Park and Suruchi Sangha. Routing is real - haversine distances between actual coordinates, a nearest-neighbour solve with crowd-aware time shifting. Pandal names, zones, metro stations and walking distances are real.

What is simulated

The festival uplift across Panchami to Dashami. The forecast ranking between pandals - Google normalises every venue to its own peak, so its numbers cannot say whether Ekdalia is busier than Deshapriya, and only the live queue reports, which are in minutes, can. The thirteen pandals with no Google coverage are modelled whenever nobody has reported them. Google's curve is an ordinary week, not Pujo. The assistant is templated, not a language model. The sponsor-exposure block in Committee Ops is a roadmap mock and is labelled as one.

Production stack

TodayLive reports on Workers + D1, Google Popular Times for shape
CountingYOLO-class person detection on existing CCTV, edge boxes at gates
ForecastGradient-boosted model on last year's counts + metro + weather
AssistantClaude with the crowd model exposed as tools, Bangla-native
RouteOR-Tools VRP with time windows, GTFS for metro legs

Pilot scope

One season, four committees, four gates each. Fixed fee per committee, hardware at cost. Deliverable is the ops dashboard plus a post-season footfall report the committee can hand to its sponsors — which is the artefact that renews the contract.

Why now

Cameras are already installed at every major pandal for security. Vision models got cheap enough to run counting on a ₹25,000 edge box. And the Bangla-language barrier that killed earlier consumer attempts is no longer a barrier.

The moat

Year one is a product. Year three is the only historical footfall dataset for the largest public gathering in the country — which is what the tourism department, the brands and the insurers all need and nobody else will have bothered to collect.