A surgeon uploads a cataract operation. OcuFlow returns a timeline segmented into the 13 canonical surgical phases — from a real, trained deep-learning model (0.918 validation accuracy, above the published Cataract-1K baselines) — then layers on motion analysis of the instruments to answer the question a trainee actually asks: "What did a better surgeon do differently, and am I getting better?"
Under the hood: a ResNet50 + MS-TCN phase model (the published TeCNO recipe, trained on Cataract-1K), hand-labelled instrument tracking at 60 fps via an in-house web labeller with Lucas-Kanade propagation, and per-phase economy-of-movement metrics grounded in the clinical motion-analysis literature — wrapped in a deployed Next.js + FastAPI product with a self-contained investor demo.
Phase recognition, 4-fold CV — beats the paper's 0.78–0.85
All ML trained on a MacBook M2 Pro
Hand-labelled instrument track, 100% coverage