Stas Kulesh

I built a robot that scans Kraków’s streets for graffiti using Mapillary’s imagery. It processes 3,500 frames across eight districts by: - Fetching data: Pulling Kraków images and GPS/timestamps via Mapillary’s Graph API. - Segmenting: Running Meta’s SAM 3 model on Apple Silicon via MLX to identify graffiti candidates. - Classifying: Using CLIP zero-shot to filter out non-graffiti objects like signs and trees. - Geocoding: Converting detections into Polish addresses and Straż Miejska districts via ArcGIS. - Filtering: Retaining only the newest capture per wall. - Reporting: Preparing payloads for Kraków’s city council online server The tool currently operates in dry-run mode, but could automate real reports to the city. So far, it has identified 471 graffiti sites across eight districts, showing that existing city infrastructure can be used to pre-fill reporting forms at scale. If you work in gov-tech, would you prefer to integrate this into existing reporting workflows or use it as an analytics layer?