Lewis Tunstall

Tunstall holds a theoretical-physics PhD from the University of Adelaide. Fulbright’s 2010 record identifies his earlier physics honours degree and doctoral research, and Bern’s course archive lists him teaching machine learning for physicists in 2020. His older owned biography describes enterprise NLP, time-series and topological-data work, including giotto-tda and anomaly detection. At Hugging Face he has worked on open learning tools and coauthored Natural Language Processing with Transformers.[4][5][6][2]

Teach a small model a repeatable proof-writing process

QED-Nano starts with supervised training on proof examples, then uses reinforcement learning with rubric-based rewards. A reasoning cache adds iterative summarize-and-refine cycles so the model can work through long proofs across steps. The coauthored study releases its four-billion-parameter model, datasets and training/evaluation pipeline. Its contribution is an open recipe for studying proof generation, with reported performance tied to the paper’s problems and scoring process.[7]

Make geometric features usable in ML pipelines

giotto-tda brings topological data analysis into a scikit-learn-compatible Python interface. Preprocessing, feature extraction and visualization can operate on several forms of data, with optimized implementations beneath the API. Its modular interface lets researchers combine geometric representations with ordinary machine-learning workflows instead of maintaining separate experimental code for every dataset. Tunstall coauthored the toolkit paper and described developing and maintaining the library.[9][2]