I built CoachAI because of a gap I kept watching in gyms: people show up consistently, train hard, and still get hurt or plateau — not because they lack discipline, but because nobody is watching how they move. trainers are expensive or don't fit into the schedule. Most people just guess.
So the core question became: can a phone actually see a bad rep?
Turns out yes. CoachAI uses the iPhone camera to track your movement in real time and tell you when your form breaks down — all on-device, nothing uploaded. Around it sits a program engine that adjusts properly instead of shuffling random exercises.
The biggest shift in my process was order of operations. My original onboarding asked for goals, height, weight, injuries — five screens before you ever saw the thing that makes the app different. People bounced. I've since flipped it: you do one tracked rep first, then we ask questions. Show the magic, earn the form fill.
On process: I shipped this before I was comfortable. There are rough edges I'd normally have spent another three months sanding down in private. But a polished app nobody has used is just a very expensive opinion. I'd rather have real people telling me what's wrong than keep guessing on my own.
Still early, still rough in places. If you try it, I want the harsh feedback — especially on the tracking accuracy.
Try it and see if it catches your form breaking down. That's the part I'm proudest of, and the part I want people to try out.
iOS 17+, free to start: https://apps.apple.com/app/coach...
Report
I love the idea. I am curious if you have thought about using this for other sports? I recently started playing tennis again and would love something that could analyze my swing and give me feedback. The setup my be a little tricky but you could have airpods in and get realtime feedback.
@elliot_myhre you're more right than you know about the audio part. our capture path is rear-camera, which means the screen is facing away from you mid-set — so coaching is already voice-only by necessity. airpods is genuinely how people use it.
Currently we don't support sports but our engine is built to handle it just the same. labeling your swings, calculating your speed. but the model will require a bit of training on tennis/other sports data. it's in the pipeline for once we have a more solid customer base.
Report
congrats on the launch! I'm wondering what the different applications of this tool are or if it can analyze all types of movements.
rep counting works across a decent spread of common lifts — it's driven by joint-angle range-of-motion, so anything with a clean repeating angle signal (presses, rows, curls, squat patterns) gets counted, and static holds like planks get timed instead.
form analysis today that's two cues — knees caving in, and back rounding. that's it. depth, tempo, neck position, hip sag are in the upcoming release.
We're currently supporting 18 exercises with more to come in the upcoming releases, but those are just enough to cover most everyday workouts.
respect for shipping this rough and asking for harsh feedback. the part I'd worry about is camera angle on the heavier compound lifts - a phone propped up to one side can miss depth or spinal rounding on a squat or deadlift that a trainer standing at the right angle would catch immediately. if the app gives a thumbs-up on form when the angle just didn't have visibility into the thing that actually matters, that's arguably worse than no feedback at all, since people will trust it and load up more. is there anything that tells the user when the angle itself isn't good enough to judge a given movement?
@omri_ben_shoham1 this is the right thing to worry about, and the honest answer is that we're only partway there.
two things work in our favour. we're on ARKit body tracking, so we get 3D joint positions, not 2D keypoints off a flat image — depth is measured rather than guessed from silhouette, which makes a 45° phone placement much less fatal than it would be for a 2D pose model. and the app never gives a thumbs-up. there's no "good rep" state. it only speaks when it detects a fault.
but that second thing is also exactly the bug you just found. silence currently means either "that was fine" or "I couldn't see it well enough to tell" — and the user can't tell those apart. same failure mode you're describing, just quieter. we are dealing with this by calling out every trusted rep.
what exists today: a watchdog that notices when you're clearly moving but no reps are registering, and hands you manual entry instead of silently eating the set. what doesn't exist: a per-joint confidence gate that says "I can't judge your spine from this angle, move the phone." that's the next thing on the list.
The real-time form correction is what makes this more than a rep counter. I work on AI for older adults, and camera-based movement analysis is exactly what home physical therapy and fall prevention has been missing. How sensitive is CoachAI to camera angle and framing? In my experience that is the thing that quietly breaks pose estimation once people are not in a clean gym setup. Have you thought about a rehab or senior mobility mode down the line?
@igorgurovichthe angle question is the one I get from everyone who's actually shipped pose estimation, so — properly:
we're on ARKit body tracking rather than a 2D keypoint model, so we get 3D joint positions and a lot of the classic "user stood at 45° and the skeleton folded" failures don't hit us the same way. what does still hurt: occlusion, low light, and anything seated or partially framed.
the other thing worth knowing for your world: it's all on-device. pose data doesn't leave the phone.
rehab and senior mobility I've deliberately stayed out of — different regulatory posture, different validation bar, and the failure cost is higher than in a gym. what I'd be open to is the supervised end of it: programs built for older adults with a human still in the loop, the app doing measurement rather than clinical judgement. if that's a line you think is real, I'd like to hear where you'd draw it.
Replies
CoachAI
I love the idea. I am curious if you have thought about using this for other sports? I recently started playing tennis again and would love something that could analyze my swing and give me feedback. The setup my be a little tricky but you could have airpods in and get realtime feedback.
CoachAI
@elliot_myhre you're more right than you know about the audio part. our capture path is rear-camera, which means the screen is facing away from you mid-set — so coaching is already voice-only by necessity. airpods is genuinely how people use it.
Currently we don't support sports but our engine is built to handle it just the same. labeling your swings, calculating your speed. but the model will require a bit of training on tennis/other sports data. it's in the pipeline for once we have a more solid customer base.
congrats on the launch! I'm wondering what the different applications of this tool are or if it can analyze all types of movements.
CoachAI
@ethan_cheng thanks!
rep counting works across a decent spread of common lifts — it's driven by joint-angle range-of-motion, so anything with a clean repeating angle signal (presses, rows, curls, squat patterns) gets counted, and static holds like planks get timed instead.
form analysis today that's two cues — knees caving in, and back rounding. that's it. depth, tempo, neck position, hip sag are in the upcoming release.
We're currently supporting 18 exercises with more to come in the upcoming releases, but those are just enough to cover most everyday workouts.
Dial
respect for shipping this rough and asking for harsh feedback. the part I'd worry about is camera angle on the heavier compound lifts - a phone propped up to one side can miss depth or spinal rounding on a squat or deadlift that a trainer standing at the right angle would catch immediately. if the app gives a thumbs-up on form when the angle just didn't have visibility into the thing that actually matters, that's arguably worse than no feedback at all, since people will trust it and load up more. is there anything that tells the user when the angle itself isn't good enough to judge a given movement?
CoachAI
@omri_ben_shoham1 this is the right thing to worry about, and the honest answer is that we're only partway there.
two things work in our favour. we're on ARKit body tracking, so we get 3D joint positions, not 2D keypoints off a flat image — depth is measured rather than guessed from silhouette, which makes a 45° phone placement much less fatal than it would be for a 2D pose model. and the app never gives a thumbs-up. there's no "good rep" state. it only speaks when it detects a fault.
but that second thing is also exactly the bug you just found. silence currently means either "that was fine" or "I couldn't see it well enough to tell" — and the user can't tell those apart. same failure mode you're describing, just quieter. we are dealing with this by calling out every trusted rep.
what exists today: a watchdog that notices when you're clearly moving but no reps are registering, and hands you manual entry instead of silently eating the set. what doesn't exist: a per-joint confidence gate that says "I can't judge your spine from this angle, move the phone." that's the next thing on the list.
Refocus
The real-time form correction is what makes this more than a rep counter. I work on AI for older adults, and camera-based movement analysis is exactly what home physical therapy and fall prevention has been missing. How sensitive is CoachAI to camera angle and framing? In my experience that is the thing that quietly breaks pose estimation once people are not in a clean gym setup. Have you thought about a rehab or senior mobility mode down the line?
CoachAI
@igorgurovichthe angle question is the one I get from everyone who's actually shipped pose estimation, so — properly:
we're on ARKit body tracking rather than a 2D keypoint model, so we get 3D joint positions and a lot of the classic "user stood at 45° and the skeleton folded" failures don't hit us the same way. what does still hurt: occlusion, low light, and anything seated or partially framed.
the other thing worth knowing for your world: it's all on-device. pose data doesn't leave the phone.
rehab and senior mobility I've deliberately stayed out of — different regulatory posture, different validation bar, and the failure cost is higher than in a gym. what I'd be open to is the supervised end of it: programs built for older adults with a human still in the loop, the app doing measurement rather than clinical judgement. if that's a line you think is real, I'd like to hear where you'd draw it.