Work
Shipped systems, published research, and things built out of curiosity. Each entry carries the problem, the approach, and the number that mattered.
Synthetic Data for Redaction Detectors
Diffusion-based synthesis with VLM-guided auto-labeling — supplied 30% of all training data and cut research-to-production from 8 months to 2–3.
Multi-Object Detection and Tracking Standard
An org-wide standard for detection and tracking, adopted across three research and product teams.
CLIP Exercise Recognition
Zero-shot exercise recognition from still images and video, built on OpenAI CLIP — no labelled training set required.
Driver Awareness
A demo built on our detection and facial-analysis APIs — in-cabin attention from gaze and head pose.
License Plate Recognition
Deep LPR system improving on the legacy pipeline by 20% in production.
DAGER: Deep Age, Gender and Emotion Recognition
State-of-the-art age, gender and emotion recognition — outperforming competitors including Microsoft by a large margin.
Car Make, Model and Color Recognition
Fine-grained vehicle attribute recognition from a single frame, in traffic.
Parent-Offspring Resemblance
CVPR 2014 — determining family resemblance via gated autoencoders.
Video Face Recognition
CVPR 2013 — mean sequence sparse representation classification on movie trailers.
Face Recognition for Web-Scale Datasets
CVIU 2013 — open-universe face identification with a 100× speedup over the base method.