Machine Learning Engineer
Cerebra AI
Jan 2024 — PresentImproved stroke segmentation and classification results, and shipped a bilingual meeting task-extraction pipeline that cut processing time for a two-hour recording from 40 to 15 minutes.
• Integrated TransUNet for ischemic stroke segmentation, improving Dice from 0.43 to 0.58, and replaced the pipeline classifier with DINOv2 for a further +3%; both results drove a company-wide shift to UNet-based architectures. • Migrated segmentation from 2D to 3D architectures (nnU-Net, MONAI) for the regulatory submission, improving hemorrhagic stroke Dice from 0.78 to 0.85, now going through FDA evaluation in the USA.