
Real-time
on-device inference
The story
From problem to proof.
PulseFit came to us with a bold idea: a dumbbell workout app that watches your form like a personal trainer. Their users were training alone at home — and bad technique was leading to stalled progress and preventable injuries.
We designed and shipped an on-device pose-tracking system that analyses every rep in real time, flags form errors as they happen, and coaches users toward safer, stronger movement — all without a single frame leaving the phone.
What we brought
The challenge
Real-time form coaching, on a phone, with zero cloud latency
Form correction only works if feedback arrives during the rep — not after it. Cloud round-trips were too slow, and off-the-shelf pose models couldn't distinguish a risky lower-back arch from a legitimately deep hinge.
- Sub-100ms inference required on mid-range devices
- No training data existed for dumbbell-specific form errors
- Battery drain made always-on camera analysis impractical
What we did
Four moves that made it work.
Built the dataset
We captured and annotated thousands of reps with certified trainers labelling correct and faulty form across body types, lighting and camera angles.
Engineered the model
A pruned pose-estimation network with a temporal error-detection head — small enough to run on-device, sharp enough to catch subtle form breaks.
Designed the coaching loop
Feedback had to motivate, not nag. We tuned cues to arrive mid-set with precise, actionable language tested with real users.
Hardened the pipeline
On-device inference with graceful degradation, plus a feedback loop that lets trainers flag misclassifications for continuous retraining.
The product
See what we shipped.


What we achieved
The numbers that followed.
PulseFit launched its AI coach to a waitlist of 40,000 users. Form-related support tickets collapsed, and the app's store rating climbed within weeks of release.
<100ms
On-device inference latency
94%
Form-error detection accuracy
+38%
Workout completion rate
4.8
App store rating post-launch
“JSXDOM didn't just build a model — they built the feature our entire marketing is now built around. Users genuinely believe a trainer is watching them.”
Maya Chen
Founder & CEO, PulseFit
Someone's case study like this is being written right now. It could be yours.
Every engagement above started with one free, no-obligation conversation.