Inertia

Problem

Sprint coaching runs on a stopwatch and a guess. Ground contact time is the number that actually separates athletes at speed, and measuring it normally takes a force plate and a lab. Neither one travels to a high school track.

Architecture

Two halves that meet in the app.

Video goes to a pose pipeline. SAM 3D Body runs on RunPod serverless and returns MHR70 keypoints per frame. Ground contact events, per-step timing, and trunk lean angle come out of that keypoint stream.

The other half is hardware. Timing gates are ESP32-S3 receivers talking over ESP-NOW, paired with analog IR emitters built around a TLC555. The emitters are deliberately dumb. No microcontroller, no firmware, no pairing. They put out a modulated beam and nothing else. All of the intelligence sits in the receivers. Custom PCB with a USB-C and LiPo power path.

Athlete baselines build through Welford's online algorithm, so the biomechanical fingerprint updates on every rep without keeping the full history around.

What's interesting

Putting all the smarts on one side of the gate. An emitter that cannot be misconfigured cannot desync halfway through a session, and desync is the failure mode that kills timing systems in the field.

Welford's also changes what the numbers mean. Drift shows up against an athlete's own baseline instead of a population average, which is the comparison that actually matters to a coach.

Status

Pre-alpha, in active development. The pose pipeline and receiver firmware work. The PCB is on its second revision.