Computer Vision Engineer — On-Device Detection (Founding Contractor). Remote (US time zones preferred) | Contract | Pre-revenue, deferred cash + equity (Clone)

ABOUT US
BullzEye is a small, pre-revenue golf-technology company. Our iPhone app scans a putting green in about 30 seconds and reads the line, entirely on-device: computer vision, LiDAR depth, and a physics solver, no cloud round-trip. Patent-stage, beta targeted for late 2026. The detector is the heart of the product, and it is not good enough yet. That is the job.

WHAT YOU'LL OWN
You own ball-and-cup detection end to end, from a working-but-mediocre on-device baseline to shipping quality:

• The model lineage: find where it fails (small objects, hard lighting, wet greens, long putts, clutter) and drive measurable improvement. Making the training pipeline reproducible is part of the mandate.
• On-device performance: export to Core ML / ONNX, NMS baked in, quantized, benchmarked to run inside a real-time scan budget on a phone.
• Calibrated refusal: the product would rather decline a bad frame than give a wrong line, so you own the abstention behavior.
• The evaluation harness: per-condition slicing, held-out scoring, regression tracking across versions.
• The licensing boundary: we ship a closed-source binary, so training stays on permissive stacks (torchvision, RF-DETR, Apple Create ML).

WHAT WE'RE LOOKING FOR

• Production object-detection experience in a real domain (shipped, not notebooks).
• On-device / edge inference: Core ML, TensorFlow Lite, ONNX Runtime, TensorRT, or similar. You've made a model small and fast, not just accurate.
• Comfort with domain shift and hard negatives; real evaluation fluency (mAP, PR curves, per-slice analysis).
• You scope down: the simplest thing that measurably works.

SENSING BACKGROUND IS OPEN
Today the depth source is iPhone LiDAR, but the architecture is source-agnostic (structured light, multi-sensor, stereo, NIR on the roadmap). Stereo / structured light / ToF / LiDAR / NIR backgrounds all map. Having shipped iPhone LiDAR specifically is NOT required.

LOGISTICS
Contract to start, remote, pre-revenue with deferred cash and equity. Tell us about one detection model you took from "promising" to "actually shipped," and what was hard about the last 10 percent.