Project
PaceAI
Proceris — closed source
Turns long-form workout video into timestamped equipment control profiles for connected fitness hardware.
- React
- Node.js
- Python
- Multimodal
Problem
Connected fitness OEMs hand-program every video workout (trainer cues, pacing changes, terrain shifts, intensity zones) into device control profiles. With libraries in the thousands and roughly two hours of manual work per video, content throughput is the bottleneck. Programmers want their judgment in the loop; they just don't want to retype the same incline curve a thousand times.
Approach
A three-stage pipeline. Stage one: multimodal ingestion. Video and audio are processed in parallel: FFmpeg for frame and waveform extraction, transcription for trainer dialogue. Stage two: an LLM agent encodes exercise science domain knowledge (RPE, heart-rate zones, work-to-rest ratios, progressive overload, manufacturer incline ranges) and combines extracted cues with programmer intent (series goals, trainer style, progression notes) to draft a control profile. Stage three: a human-in-the-loop review UI where programmers approve, edit, or reject the draft before it exports into the partner OEM's production control schema.
Deployment
Built for enterprise data-residency requirements: pipeline runs in a secure cloud environment by default, with an on-prem option for partners that can't ship video outside their own network. Structured output maps directly into the partner's existing control schema.
Stack
Node.js · Python · FFmpeg · multimodal LLM pipeline · structured-output schemas · React review UI · containerized deployment