Where water is escaping
Water balance and minimum night flow per sub-division, plus streaming anomaly detection on flow and pressure, weighted by pipe and valve age.
In development: prototype, pilot-ready
JalaLens: one ranked list for leaks and lost revenue
A map-first AI layer for water utilities. It puts the pipe map, flows, billing and building footprints on one grid of hexagons (about 0.1 sq km each) and ranks where field teams should go next.
What it finds
Who is connected but not billed
Buildings within 30 m of a water main with no billed account, using open building footprints. Later, property and power-record joins, the method the board already uses.
Which bills look wrong
Zero or flat reads, sudden drops, a domestic tariff with commercial-size use, stuck meters. Each flag shows the rupees per month at stake.
What to do today
The day’s list fits team capacity. Field results feed back, so the next list is better. A lead means “inspect”, never an automatic penalty.
Where it stands today
We state our real level: technology readiness level 4 to 5. There is no customer and no field use yet.
Done
- The full pipeline runs end to end on Bengaluru’s real public network maps: 11,670 km of mains, 48,982 valves and 43 sub-divisions (KGIS, boundaries).
- One full run takes about 4 minutes on one laptop, measured with synthetic bills.
- The billing scorer, on synthetic bills with planted cases, puts 68% hits in its top 100 against 7.1% for random picks.
Not done yet
- Leak engine testing on the public BattLeDIM benchmark is in progress (a simulated town with a known leak list).
- No run yet on a utility’s own SCADA or billing data.
- No field user yet. The pilot below is the route to operational use.
How it is built
- Models and methods
- Hydraulics with EPANET and WNTR (US EPA open source); anomaly detection with PyOD and river (streaming).
- Geography and data
- GeoPandas, H3 hexagon grid, DuckDB.
- Service and screens
- FastAPI (REST, OpenAPI), Streamlit and a MapLibre web map. Packaged with Docker.
- Inputs
- GIS layers (GeoJSON or KML), SCADA exports (CSV), billing extract (CSV or Excel).
- Outputs
- Map, GeoJSON, CSV and a documented REST API.
- Data protection
- Installed on the utility’s own server or VM; consumer IDs pseudonymised; designed for India’s DPDP Act, 2023.
- GPU work planned next
- GPU work we plan next: training graph and time-series models for leak detection on simulated pipe networks (NVIDIA CUDA / PyTorch).
- Licences
- All parts MIT, BSD or Apache licensed: no licence fees and no vendor lock-in.
The pilot we propose
Two sub-divisions chosen with the utility, 60 days of daily use, then handover.
- Weeks 1 to 2: read-only data access under an NDA, install inside the utility, first run on real data.
- Weeks 3 to 8: daily ranked lists; about 200 leads checked on the ground; results fed back.
- Day 60: a report on hit rate against random visits, confirmed illegal connections and billing corrections, leaks found and time to repair, and night-flow change.
Run a 60-day pilot with us
Two sub-divisions, your own data, measured against random field visits.