All work
Computer vision
Parcelworks
2025
An on-device vision model that counts warehouse stock from a handheld scanner — no network, no cloud round-trip, no waiting.
Scroll to resolve
18 ms
Per frame, on device
Client
Parcelworks
Role
ML engineer, vision
Stack
PyTorch · YOLO distillation · CoreML · Rust
Year
2025
The problem
Stock counts happened once a quarter because the existing tool needed Wi‑Fi that most aisles didn’t have.
What I built
I distilled a large detection model into a 9 MB student that runs on the scanner’s own chip, then wrote a Rust post-processor that merges overlapping frames into a single count.
The result
Counts now run nightly on every aisle at 18 ms per frame, and shrinkage is spotted weeks earlier.