AIT City Intelligence Lab, Vienna · Sep 2025 – Jul 2026
AI Engineering Internship (Sep–Dec 2025) → Master's Thesis, IAAC × AIT (Jan–Jul 2026)
Problem
Cities allocate curbside space — parking, deliveries, transit stops, pedestrian zones — using static zoning that does not reflect how streets behave over the day, week, and season. The challenge was to turn raw parking events, pedestrian counts, weather, land-use data, and street geometry into a single decision pipeline that identifies where curbside intensification is justified and what kind of intervention fits each street.
My Contribution
I designed and built the full pipeline end-to-end: automated data ingestion, sensor snapping, temporal feature engineering, graph construction, parking reconstruction, pedestrian imputation, weekly profile aggregation, GMM clustering, MultiGCN training, counterfactual scenario analysis, and frontend export. I also translated the model outputs into street-level archetypes, flexibility windows, and intervention scores that can be read by planners without ML context.
Technical Approach
Data ingestion: A 12-step pipeline pulls parking occupancy, pedestrian counts, Open-Meteo weather, and Melbourne Open Data CLUE datasets into a shared framework. The raw inputs are aligned to 1,397 eligible arterial and council-major streets over 14,400 fifteen-minute bins.
Feature engineering: Parking and pedestrian streams are reconstructed into complete street-level time series, then compressed into weekly behavioural profiles with 84 temporal features plus engineered scores for morning, midday, evening, weekend ratio, and parking flexibility. Spatial context is encoded through CLUE land-use aggregation and a k-NN street graph.
Modelling decisions: GMM clustering was used instead of fixed categories so the street archetypes could emerge from the data itself. Pedestrian gaps are filled with XGBoost using static, temporal, weather, and spatial-lag features, then a MultiGCN learns next-step pedestrian flow from both spatial and semantic graphs. Counterfactual simulation and opportunity scoring turn the model outputs into intervention priorities.
Automation: The full pipeline — from raw fetches through export-ready GeoJSON and summary JSON files — is automated for reproducible runs, with validation at each stage and CI/CD support via GitHub Actions.
Thesis extension: The thesis work extends the original internship pipeline with longer temporal coverage, graph-informed imputation, counterfactual intervention simulation, and updated model interpretation for a publication-ready system.
Stack
Python · Pandas · NumPy · scikit-learn · XGBoost · PyTorch · GeoPandas · GIS libraries · Flask · GitHub Actions · SQL · Git · Bash
Result
A production-ready curbside intensification pipeline that outputs street archetypes, flexibility windows, intervention scores, counterfactual scenario results, and frontend-ready exports for the Melbourne CBD map. The system is built around a 1,397-street graph, produces complete 14,400-bin street histories, and is ready for decision support and further deployment work.