Identified 11 vulnerabilities in the swiss voting advice application Smartvote with some allowing for more than 3.5x visibility gains for individual parties. Proposed 10 mitigations to significantly reduce or eliminate these vulnerabilities. Findings are being adopted in Smartvote's redesign for the next elections.

Ranked 4th out of 12 teams in the small track of the DataComp Challenge, an ML benchmark where the goal was to filter a CommonCrawl image-text dataset to train a CLIP model evaluated on 38 zero-shot downstream tasks, using a combination of cross-modality filtering and content alignment.

Developed an interactive dashboard to predict basketball game outcomes based on in-game stats and explain predictions using SHAP values. Users could modify team statistics to explore what-if scenarios.

Core contributor to the open-source time series forecasting library Darts by Unit8. Optimized the most popular regression forecasting models by vectorizing computations achieving a speedup of up to 400x.

Developed a model that enables constant-time approximate shortest path distance queries on road networks, achieving an average mean relative error of less than 10%.