Work

Seattle, WA — Owned the full experimentation pipeline for acoustic screening on Google HeAR embeddings + PyTorch classifier: 0.8245 AUC (78.7% sensitivity, 71.3% specificity). Built a robustness pipeline (event isolation, codec/noise/gain augmentation) lifting AUC 0.782 → 0.8245, and shipped inference via a Flask API.

Remote — Built production AI systems combining retrieval + generation over 1,400+ biomedical manuscripts, and automated research bots with Gemini File Search and the ElizaOS agent framework.

Remote — Cut accelerator memory for a large 3D nnU-Net segmentation model by redesigning multi-channel inputs as single-channel, improving GPU training efficiency.