Selected work
Products I built from the first commit, and the engineering behind them.
A physical AI data marketplace and a complete training platform, built from an empty repository. 3D tools that stay smooth with millions of points, images too large for a browser to open, an autonomous labeling agent, and a build migration that took a whole team from a minute of waiting to under a second.
Deepen Refinery
The marketplace for physical AI data, built from the first commit to launch
Robotics, self-driving and medical teams come here for the real-world data they train on. I shaped the product and built the entire frontend: three workspaces, every flow, and sample viewers that open a dataset at full size.
Learning Hub
A training platform I built end to end, from product plan to grading engine
Courses, quizzes, hands-on practice in the real annotation editor, automatic grading, and feedback that draws every mistake onto the data in 2D and 3D. I planned, designed and built all of it, frontend and backend.
Multi-camera view
Every camera on screen at once, tied to the 3D scene
Labelers were flipping between cameras several times a frame, for hours a day. I designed and built a multi-camera view for a dense 3D editor: a resizable grid, spatial cues linking each image to the lidar scene, and auto-zoom that finds the object in every camera.
Gigapixel annotation
Gigapixel annotation: 100,000-pixel images that open and pan like a map
A single image could need 40 GB of memory just to decode. I designed the approach, built the tile-pyramid backend, and wired it into the annotation editor, so every image label type works at that scale.
Webpack 5 to Rspack
From a minute to under a second: moving a huge frontend to Rspack
I initiated and delivered the move from Webpack 5 to Rspack for a frontend of several thousand components. Dev startup went from about a minute to under a second, and cloud builds from about six minutes to under one, roughly 85% cheaper.
Labeling Agent
An autonomous labeling agent: orchestration, vision pipeline and benchmark
An AI agent that coordinates vision-language models and specialist vision models to generate, refine and review labels. On real autonomous-vehicle data it took mask IoU from 0.2–0.4 to around 0.95. I conceived it and built all of it: the orchestration, the vision pipeline, and the benchmark that scores every technique.
Dense 3D performance
Up to 55× faster rendering in million-point 3D scenes
The 3D labeling editor stuttered in dense lidar scenes and crept up on memory through long sessions. I profiled it, found the causes and fixed them: label text about 55× faster, 3D boxes about 15×, dense-scene interaction about 2.5×.
Deepen Validate
Deepen Validate: a data-quality product I defined from the first idea
Configurable rules — predefined, manual, and AI rules written in plain language — run by a validation engine over every label and attribute in a dataset, at any scope. I defined the product experience end to end and built the entire frontend.
PII anonymization
A privacy pipeline that finds faces and license plates where models usually fail
I evaluated open-source models like RetinaFace and EgoBlur, trained my own, built the ground truth to judge them, and developed a detection pipeline that holds up in motion blur, aerial views and low light. Then I built the tool around it.
Video-language annotation
A complete video-language annotation tool, designed and built across every flow
A tool for building video-language datasets: synchronized video streams, AI-drafted descriptions, and timed responses people can correct. I owned the UX, the design and the entire frontend.
Text and audio editor
A new annotation editor for text, audio, and both together
A labeling editor built new, beside the 2D and 3D editors, with three modes on one architecture: overlapping text annotations, a custom audio timeline, and scenes mixing several of each.