FitPulse
AI personal trainer in your pocket
200K+
App downloads
91%
Form accuracy
28 min
Avg. session length
42%
D30 retention
The Challenge
A Canadian fitness startup wanted to democratise personal training by using the phone camera to analyse exercise form in real time — without requiring a GPU server for inference.
The Solution
On-device MediaPipe pose estimation + TensorFlow Lite form-scoring model (runs at 30fps on mid-range phones). Backend generates personalised plans using workout history. No cloud inference needed.
How We Built It
Model Research
MediaPipe Pose benchmark vs BlazePose, TFLite quantisation strategy, latency budgeting.
Training Data
Labelled 8,000 exercise reps across 20 movements with certified trainers.
On-Device Model
Int8 quantised TFLite model, Flutter plugin bridging native camera feed.
Plan Engine
Periodisation algorithm generating weekly mesocycles from fitness goals and history.
Launch
TestFlight beta → App Store launch → influencer campaign delivering 50K downloads week 1.
"On-device inference was a hard technical problem — we'd been quoted 6 months by others. Ubikon shipped a working, accurate model in the Flutter app in 18 weeks total."
Marc Leblanc
Founder, FitPulse
Tech Stack
Start a similar project
Free proposal within 24 hours. Fixed-price contract. 100% source code ownership.