VBT Tracker

One-liner: A camera-first velocity-based training PWA that turns any phone into a bar-speed sensor — with AI autoregulation, BLE hardware support, and a live coaching dashboard.

Origin (the Build Loop)

Because I compete in Olympic weightlifting and train by velocity — bar speed is the honest readout of how heavy a weight actually is on a given day, I kept hitting the fact that real bar-speed data is hard to capture: commercial VBT units cost hundreds of dollars, and there was no good way to just point a camera at the bar and get velocity, so I built a camera-first VBT tracker that runs a computer-vision pipeline in the browser that reads bar velocity from video, falls back to a BLE sensor, and autoregulates the session in real time. It taught me how to ship an on-device ML vision pipeline that stays usable offline on a phone at the gym.

The problem

Velocity-based training only works if you can actually measure bar speed rep to rep — the velocity tells you when to add load, when to stop, and where fatigue is setting in. In practice that data is locked behind expensive linear-position transducers, and the software around them is clunky. I wanted the measurement to be as cheap as the camera already in my pocket, work without a connection in a gym with bad Wi-Fi, and speak the language a lifter and coach actually use: velocity zones, e1RM, RPE, and load-velocity profiles.

What it does

• Camera VBT: point a phone at the bar and get per-rep velocity from an in-browser vision pipeline (barbell detection, pose estimation, rep detection, velocity calculation).
• BLE live mode: connect a Bluetooth bar-speed sensor and mirror live velocity in real time via the Web Bluetooth API.
• Set review: edit reps, tag RPE, view a bar-path overlay, and read per-rep estimated 1RM.
• Workout tracking: multi-set logging with quick weight adjust and a live load-velocity profile as the session builds.
• Post-set autoregulation: each set is scored for zone adherence and fatigue, returning a concrete call — increase load, decrease load, maintain, or stop.
• Analytics: velocity trends over time, zone distribution, and fatigue alerts.
• Coach mode: a multi-athlete live BLE dashboard for watching a whole group’s bar speed at once.
• Session history & offline sync: sessions cache locally and sync when a connection returns, with a visible sync indicator.

How it’s built

• Stack: Vite 8 + React 19 + TypeScript 6 PWA; Tailwind CSS 4; Zustand 5 state; React Router 7; Recharts 3; TensorFlow.js 4 + MediaPipe Tasks Vision for the camera pipeline; Web Bluetooth API; IndexedDB via Dexie 4 for offline cache; vite-plugin-pwa service worker. Node.js + Express 5 backend on Neon serverless PostgreSQL. A Python FastAPI microservice handles autoregulation. nRF52840 firmware reference for the hardware path. Deployed on Replit; GitHub Actions CI pins Node 22 + Nix stable-25_05 for Replit parity.
• Notable engineering:
  • On-device vision pipeline — a modular vision/ service (BarbellDetector → PoseEstimator → RepDetector → VelocityCalculator → VisionManager) that computes bar velocity from a live camera feed, lazy-loaded so the ~1.2 MB TF.js/MediaPipe chunk (~301 kB gzip) never blocks the app shell.
  • Offline-first PWA — IndexedDB caching plus a sync endpoint means sessions logged in a dead-Wi-Fi gym persist and reconcile later; the backend DB pool initializes lazily and returns 503 on data routes until a DATABASE_URL is set, so the app boots cleanly with or without a database.
  • Three-service architecture on one Replit — PWA (5173), Express backend (3001), and FastAPI autoregulation service (8000) all start together, with the autoregulation math deliberately isolated in Python.
  • Domain-correct model — velocity zones (in-range / fast / slow), e1RM, load-velocity profiling, and RPE are first-class, driven by a single source-of-truth set of zone color tokens and utilities (velocityProcessor, zoneCalculator, oneRMCalculator).
• Architecture: camera or BLE input → client vision/velocity processing → Zustand store + IndexedDB cache → Express API on Neon PG → FastAPI autoregulation service returns per-set load recommendations, confidence, and velocity-drop % → Recharts analytics + coach dashboard.

Proof points

• 17 screen/utility components across 12 routes; two independent capture paths (camera CV and BLE hardware) feeding one data model.
• Full vision pipeline (barbell detection, pose estimation, rep detection, velocity calculation) running client-side in the browser.
• Autoregulation service returns actionable per-set calls — increase / decrease / maintain / stop — with confidence scores and velocity-drop percentage.
• Research-grounded: development priorities derived from 4 peer-reviewed VBT studies (accuracy benchmarks, rep-detection targets), documented in the repo.
• CI enforces Replit/Docker environment parity on every push.

What to show

• Demo: Deploy the Replit build to a public URL. Best live demo is camera VBT on a phone — point it at a loaded bar and show live per-rep velocity, then the post-set autoregulation call.
• Visuals needed: the CameraLiveLiftScreen mid-rep with bar-path overlay; the load-velocity profile building across a set; the coach-mode multi-athlete BLE dashboard; the analytics velocity-trend view. No sensitive data to blur.

Cross-links

• Part of the weightlifting toolchain with Mind Games (the anchor case study — the competition/data side of the sport) and WORKOUTFLOW (the programming/team side). VBT Tracker is the in-the-gym measurement layer of that trio.
• Shares the offline-first PWA + IndexedDB caching pattern with the wedding/DJ paperwork PWA work.

Case-study angle

Alex turned a $400 piece of gym hardware into a browser tab: a lifter who competes in Olympic weightlifting built the bar-speed sensor he wanted, ran the ML vision pipeline on-device, and wrapped it in coach-grade autoregulation.