13K naturally occurring questions from search engines, human–human, and human–LLM interactions. Accepted at CALM @ NeurIPS 2024 and NAACL Findings 2025.

Evaluation framework using questions that reflect real user behavior rather than curated benchmarks.

Web app with visualizations and search on Italian train delays. Open-source initiative to promote railway data transparency.

Re-implemented Pix2Seq-D from TensorFlow to PyTorch and adapted the architecture for video panoptic segmentation.

Used Causal Representation Learning to improve classifier accuracy on out-of-distribution data.

NLP analysis of Quotebank (178M rows) measuring lexical complexity differences across American politicians.

Fine-tuned Distilled GPT-2 with a Deberta reward model using Constitutional AI, biasing outputs toward clarity, correctness, and rigour for STEM students.

Generated realistic synthetic images of C. elegans using StyleGAN2-ADA for settings where biological data is scarce.

Built a probabilistic journey planner for Zurich using SBB data, matching Google Maps results with added confidence scores.

Survey on Nash equilibria approximation via CFR and reinforcement learning from self-play for poker bots.

Compared MCMC algorithms (Metropolis-Hastings, Houdayer) for community detection, analyzing convergence and cluster connectivity.