Bryce Evans

Cornell University

Verily

Sr. Machine Learning Software Engineer, Surgical Devices

Designed and built ML infrastructure for real-time surgical devices.
Agile team focused on early stage projects and rapid deployment of new features to clinical trails (1-3 month eng. time).

• Used TensorRT, CUDA to reduce runtimes 22%-40% in several models for cholecystectomy (gallbladder removal) and colonoscopy.
• Developed benchmarks and model understanding tools to de-risk deployment and catch issues early on.
  e.g. model extremely dependent on certain (less significant) features; many false positives when out of body.
• Took ownership in product development, prioritizing features and defining new metrics to dramatically improve surgeon satisfaction.
• Co-authored MICCAI 2022 paper on handling noisy annotations in surgical data.

Cruise

Machine Learning Engineer, Vision

End to end model training, optimization, validation, and deployment.

• Trained and validated a state-of-the-art multi-modal model for long-range 3D object detection.
• Increased mAP 24% over production 2D detection model baseline.
• Additional heads consolidated multiple production models for easier maintenance and lower overall latency.

End-to-End Perception

• Identified and documented bottleneck areas downstream that limited Perception impact on end-to-end driving.
• Developed system forwarding statistics and confidences. Class-label error reduced 45%, Distant object recall up 30%.
• Retrained Prediction models on updated Perception model outputs to solve four separate safety scenarios.

Zoox

Research Engineer, Vision

Object Re-ID and Production Vision Infrastructure

• Trained, validated, deployed a multi-class embedding model for detection association, replacing heuristics.
• Increased mAP by over 20% to greatly improve confidence tracking pedestrians and enable night driving in busy scenes.
• Sole implementor of two web-based video annotation tools. Managed 180 annotators to label millions of images for multiple tasks.
• Re-built on-vehicle image pipeline with AVX, TBB, and HEVC for 2x throughput; decreased file size by 40%.

Facebook Research

Software Engineer, Computational Photography

Video Capture for VR Applications

• Primary implementer of exploratory work for Stereo-360 Capture, resulting in a Siggraph paper and patent.

Journals & Conferences

• Leifman, George, Tomer Golany, Bryce Evans, and Ehud Rivlin. "Critical View of Safety Estimation based on Noisy Annotations." In: Medical Image Computing and Computer Assisted Intervention (MICCAI). (2022).
• Matzen, Kevin, Michael F. Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. "Low-cost 360 stereo photography and video capture." ACM Transactions on Graphics (TOG) 36, no. 4 (2017): 1-12.

Patents

• "Improved Camera Based 3D Object Detection Using High Resolution Maps" (US20220414387)
• "Image Embedding for Object Tracking" (US11430225)
• "Stereoscopic Image Capture" (US20180007344)
• "Discovering visited travel destinations from a set of digital images." (US9418482)