Hackathon: Urban Insights

Together with my good friend Anastasios Stamoulakatos, data scientist at Satalia, we took part in the City of Athens Apps4Athens 2025 hackathon (2.0 AI Edition). The hackathon concentrated on digital services to improve quality of life in Athens, Greece.

Starting with my masters thesis on Mapping Urban Problems as inspiration, and seeing how far the AI technology had advanced since, we wanted to hit the next level of an urban vision prototype.

The challenge being addressed was the lack of efficiency in current infrastructure monitoring practices by the city, as well as lack of effectiveness seen in the resulting speed of interventions in the city fabric. Athens remains, in 2025, a city with significantly reduced accessibility, and a sprawling network of decaying infrastructure. There is currently no steady monitoring of the city, while maintenance decisions are subjective and sporadic.

Our prototype focused on a data pipeline to easily go from street-view images of the urban environment, potentially collected from municipal vehicles, to maps and weekly reports of the density of urban infrastructure problems in the city (e.g. broken or narrow sidewalks, lack of green, exposed garbage, illegal parking). Such pipeline would not only help prioritize highly problematic neighbourhoods, but also analyze the time density of recurring problems, and publicise the city’s efforts in a visual evidence-based way.

Preliminary GIS preparation of an Athens street geolayer from OpenStreetMap helped create segments to base our analytic results on. Supabase was used as the primary database because of its advanced postgres geographic functionality. Next.js was used to build a basic React application with a Mapbox map (OSM compatible so street geometry matched). An ETL pipeline was built in n8n to download prefiltered street image (georeferenced) datasets from mapillary, analyze them using a multimodal LLM (OpenAI GPT-4.1-mini), and classify them against a set scorecard based on different urban quality parameters. Short textual descriptions were also generated from the images, that could later be used in detailed summaries too.

The classification results were joined to specific street segments using a mapmatching algorithm taking into account the whole dataset capture path, in order to avoid GPS outliers. Once data points are processed, map visualization and easy retrieval were carried out in the app. Aggregation by zone and time period, average scores by parameter, and easy spatial identification of the city’s performance. Performance and usability measures included: throttling and macrotasks while the map updated, hover optimisation, and segment buffers so thin lines were easy to click or tap.

We won third prize for our deliverable, including a monetary sum. We were one of the few teams with just two members, out of more than 100 contestants.