Role: Director, Computer Vision
Directed the computer-vision team building the perception stack for autonomous farm equipment. Farms are an unusually friendly setting for autonomy — repeatable routes, no pedestrians — and an unusually hostile one for the sensors: dust, glare, mud, crop occlusion, variable terrain. The goal: a tractor that can stay in lane, avoid obstacles, and finish a row without an operator in the cab.
- Balanced individual contributions (60%) and management responsibilities (40%) while leading a team of 5 computer vision and data scientists.
Perception stack
A single multi-head deep network handles the downstream perception asks — object detection, semantic segmentation, and a few task-specific heads — sharing a backbone so the embedded compute budget gets spent on representation, not on running three separate models. Trained end-to-end on the team's curated farm imagery.
3D scene understanding and localization
Beyond per-frame detection, the vehicle needs to know where obstacles are in metric space and where it is in the field. The team built 3D scene-understanding models that recover obstacle geometry and a visual localization pipeline that anchors the tractor to a previously-mapped field without depending on a clean GPS fix.
Edge deployment
Models had to run on the embedded compute already on the vehicle, inside a thermal and power envelope set by the tractor — not in a datacenter. Most of the optimization work was quantization, pruning, and operator fusion to fit the multi-head model into the available inference budget without sacrificing detection recall.
Data platform
A labeling and curation pipeline so the team could turn millions of raw vehicle log frames into a usable training set without drowning a small annotation budget. Active sampling pulled the rare cases (unusual obstacles, edge weather, dust events) into the labeling queue first.
Technical Highlights:
- Deep learning models involving multiple heads for tasks such as object detection and segmentation to name a few.
- Optimized models on embedded platforms
- Developed 3D Scene understanding and localization algorithms.
- Data platform for labeling and curating millions of images.
