Projects I've worked on, ranging from research code to side experiments.
Integrated Research Environment (IRE)
June 2026Desktop App, Rust, AI Agents
A local-first desktop environment for researchers that keeps code, literature, experiments, and AI collaboration seamlessly connected. Features a persistent Git-tracked wiki, non-blocking experiment tracking, and a native MCP bridge for agentic IDE integration.
Autoresearch HFT
May 2026Quantitative Finance, Machine Learning, AI Agents
A weekend challenge to trade six months of nanosecond-precision HFT order book data. Using a modular, agent-driven research harness, self-looping agents tested 30 strategies in under 48 hours, going from losing money to a fully automated strategy earning over €25k/month.
Multi-Modal Real Estate Search
April 2026Python, AWS, MCP
As our Datathon 2026 submission, we built a multi-stage, multi-modal search pipeline that transforms natural language queries into ranked real estate listings for the Swiss market. The result is an MCP server fully deployed on AWS enabling any compatible LLM to query real-time listings.
Diffusion Language Models Are Natively Length-Aware
March 2026Natural Language Processing, Language Modelling
Improved computational efficiency of an 8B diffusion language model by predicting output sequence length before generation, dynamically cropping the context window to cut computation with minimal performance impact.
Knowledge Graph of the Italian Law
August 2025Python, Neo4j, Knowledge Graph
During my internship at Lexroom.ai, I designed and developed a proprietary ETL pipeline to construct a knowledge graph from over 5,000 unstructured Italian Supreme Court's rulings. Implemented an out-of-core, efficient PageRank-based algorithm to assign relevance scores to graph nodes.
Do AI and Human Experts See Pneumonia the Same Way?
May 2025Computer Vision, Explainable AI, Medical Imaging
We compared CNNs with experienced professionals on pediatric chest X-rays to understand if deep learning models identify pneumonia using the same visual cues as doctors. The answer is no, and high accuracy doesn't guarantee human-interpretable reasoning.
Wordle Solver
February 2025Natural Language Processing
Optimal strategy solver for Wordle that achieves 100% success rate with an average of 2.89 guesses in 0.5 seconds.
BSc Thesis
July 2024Bioinformatics, Machine Learning
Enhanced scGPT transformer model for gene expression prediction by incorporating prior biological knowledge of transcriptional regulation.
Daily Spin Bot
December 2023Python, Browser Automation
Automated browser bot for maximizing rewards in SNAI Daily Spin game, available as executable.