TEDDY
A transformer trained from scratch to forecast first-occurrence diagnoses from longitudinal clinical event histories.
150M+raw clinical events tokenized
Selected projects
I build machine-learning systems from the model outward: data and training, evaluation and failure analysis, backend architecture, and deployment.
A transformer trained from scratch to forecast first-occurrence diagnoses from longitudinal clinical event histories.
A production-oriented mentor search system where evaluation led from an unstable LLM judge to a fine-tuned cross-encoder reranker.
Profiling and redesigning a scientific Python bottleneck reduced a large single-cell workload from more than 12 hours to minutes.
Semi-supervised image and video pipelines for lesion detection, with active learning and temporal propagation used to extend limited labels.