The data about a place, seen together.

Sentinel-1, Sentinel-2, Sentinel-3 and Sentinel-5P, through the Copernicus Data Space Ecosystem. In the code the collections are named exactly `sentinel-1-grd`, `sentinel-2-l2a`, `sentinel-3-slstr` and `sentinel-5p-l2`. We have no satellites of our own and we do not buy high-resolution commercial imagery — if you need 30 cm, that is a different conversation and a different supplier.

No, and nobody can give you that with Sentinel. Sentinel-2 passes over the same place about every 5 days, Sentinel-1 every 6–12 days, and the product needs processing time. That is why the default window for the layers is the last 7 days, ending yesterday. What you get is the best scene available in the interval you asked for, with its date stated.

The optical view is lost. We filter out scenes above 40% cloud and mask the pixels the SCL band marks as cloud, shadow, cirrus or snow — they come back transparent, not invented. When optical cannot, Sentinel-1 remains: radar passes through cloud and works at night, at the price of being read differently — it sees roughness and moisture, not colour.

No. That takes InSAR: pairs of scenes, precise orbits, a DEM and phase unwrapping. In the code this product is listed with the status `requires_pair_processing` — declared as not implemented, rather than hidden. What we do instead: screening. There is a landslide triage module that introduces itself as `flowmind-landslide-screening-v0`, heuristic, with the warning written into the API response that it is not confirmed detection.

Few, and you are better off hearing it from us. An OSINT survey of our own from July 2026 found a ceiling of roughly 34 public road cameras live in the whole country — 12 of them ANDSA, plus 22 at the border. At today's check, out of 116 streams probed, 13 answered, all of them refreshed JPEG stills, zero video feeds. The rest are declared with the reason: expired subscription, rotated token, folder gone, port unreachable. If somebody promises you 150 cameras in Moldova, ask them for the list.

No, and in its present form it could not. The 245 traffic lights are mapped as points in the data, from OpenStreetMap — the platform sees them on the map, it does not command them. Active control would mean integration with the physical controllers and liability for safety, which is a separate project, with other partners.

No. There is no trained model in the repository and no machine-learning code is called by the application; the `ml/gnn` and `ml/forecasting` directories are empty. The RL and GNN architecture in ARCHITECTURE.md is the target, marked as such in that very file.

A standard Copernicus account comes with 30,000 processing units and 30,000 requests a month, free of charge. Every tile requested draws on them. The platform cuts consumption through caching (30 minutes in memory, 24 hours in the response) and through limits on the size and the interval of a request, but it has no quota counter — if the installation carries heavy public traffic, metering and a paid plan are things to decide beforehand, not afterwards.