Sri Lanka has no live transit data. There is no GTFS feed for the country, the National Transport Commission publishes its more than 1,000 bus routes as PDFs, and Google Maps renders the route line a bus is supposed to follow with no information about actual vehicles. The standard fix, GPS units on the fleet, does not fit the fleet’s structure: over 65,000 vehicles split between the state-run SLTB and thousands of small private operators, many of whom own a single bus. No party exists that could fund, install, and maintain hardware at that scale. Mansariya uses the sensors already on board: the passengers’ phones.
The contribution model
A rider taps “I’m on a bus” and the app posts the phone’s location every five seconds for the duration of the ride. That is the entire contribution. In return, other users on the same route see the inferred bus position move across their map, live, and the roles reverse on their next ride.
A raw GPS trace from one phone does not identify a road, a route, or a vehicle on its own. The backend infers those in stages.
The inference pipeline
Ingestion is minimal by design. The phone POSTs a batch of positions, the server appends it to a Redis Stream, and the request completes. All processing runs asynchronously off the stream, so a spike in contributors does not slow the ingestion path.
The first stage snaps each trace to the road network using Valhalla’s map-matching engine, self-hosted. A commercial per-request map-matching API was ruled out because its cost scales with exactly the usage the project hopes to attract, which does not fit the project’s economics. The matched trace is then compared against all 1,000+ route polylines through an R-tree spatial index to infer which route the phone is riding.
Clustering resolves phones into vehicles. Several devices moving together along the same route are treated as one bus: DBSCAN groups the co-moving devices into a virtual vehicle and fuses their readings into a single position. One contributor yields a usable position, a second improves accuracy, and at three the position is marked verified. The fused position is published through Redis Pub/Sub to WebSocket connections, and every open map in the app updates.
The whole backend runs as a single Go binary backed by PostgreSQL with PostGIS on a Hetzner server that costs about $9 a month. Map tiles come from OpenFreeMap rendered with MapLibre, so the recurring map cost is zero and there is no metered API key with a quota to exceed.
Privacy as a structural property
The system’s input is the continuous location of commuters, so the privacy design has to hold against the server operator as well as against third parties. Mansariya is built so that the operator cannot reconstruct an individual’s travel history, as a property of the architecture rather than a policy. There are no user accounts and no login. The device identifier rotates every 24 hours, so contributions from different days cannot be linked to each other. Raw GPS data is discarded within ten minutes of arrival, and only aggregated, anonymous data persists. Storing personal traces is available strictly as an opt-in. This removes the trust question instead of answering it: the data required to identify anyone is deleted before it could be misused.
The client is React Native on the bare workflow, Android first, since Android is the dominant platform among bus riders in Sri Lanka. Route names, stop names, and every UI string exist in Sinhala, Tamil, and English, and all 1,000+ routes are cached for offline browsing.
The cold-start problem
Crowdsourced coverage starts at zero. The map shows live positions only where someone is currently riding with the app open, and on launch day that is nobody. Two mitigations lower the threshold. A single contributor produces a usable position, so one rider is enough to make a route useful, and offline route browsing with trilingual search gives the app value before any live data exists. Adoption itself is not an engineering problem, and the code does not solve it.
Other open items are narrower. ETA estimates depend on historical speed data that has to accumulate before the estimates mean much. iOS support is deferred. The clustering accuracy has not yet been validated against dense Colombo traffic, where two buses queued closely on the same route could register as co-moving devices.
Status and moving forward
The MVP is under active development, built test-first, because a spatial inference pipeline produces plausible-looking wrong output when untested. The design principle that holds throughout is to size the system to its economics: self-hosted map matching, free map tiles, and a single binary keep the total infrastructure cost at about $9 a month.
If the crowdsourcing model produces coverage, Sri Lanka gets live transit data for the price of one small server. If it does not, the cost of finding out is bounded at $9 a month, and the results will be documented publicly either way.