ML, with biology and health in mind.
Compute-Aware Automated Protein-Model Optimization
An agentic system for end-to-end protein-model optimization, evaluated through three protein-engineering tasks and ablation studies. The project explores how to make model optimization effective under compute constraints.
Supervised by T. Bikias and M.-A. Hartley
Thesis grade 6/6Best Poster Award at OxML 2026 Health & Bio Track and accepted to NeurIPS AgenticLS workshop
Built a retrieval pipeline over an immunology corpus using an entity graph and Personalized PageRank, and benchmarked it against production RAG.
Developed an 8B parameter modular multimodal medical language model trained from scratch, exploring mixture-of-experts and embedding-projection fusion.
Explored log-probability rewards with GRPO to elicit language-model reasoning across mathematics and poetry tasks.
In this 10-week project, our group of three aimed to improve a small language model on scientific question answering using EPFL exam questions. We fine-tuned and quantized the model for single-GPU training, applied Direct Preference Optimization (DPO), and expanded its knowledge with Retrieval-Augmented Generation (RAG).
Investigated multimodal large language models for food recognition.
Supervised by Prof. Marcel Salathé
Thesis grade 6/6
Fine-tuned sequence-only models to predict germline mutation pathogenicity across 70 ClinVar cancer genes, achieving an AUC of 0.875.
Supervised by Jeremy Wohlwend