Align

IOS App Screenshots

Inspiration

Low back pain is the world’s leading cause of disability; about 619 million people, per WHO. Most of us will feel it. It usually isn’t one stiff night. It’s years of sitting, stacked. Bad compound interest.

Catch it early and the same math runs the other way: small changes while they’re still cheap. That’s Align. A preemptive check-in, not a posture fixer. Most apps skip that and flash a score from one photo. We continuously collect data and improve our model with every new submission.

What it does

Put an iPhone on a stand. Level it. Take four photos: front, right, back, left.

Tap Analyze posture. Pose runs on those stills. Recap shows projected angles when landmarks are good, and says so when they aren’t. Four photos finished is not “good posture.”

Talking is optional. Ask out loud if you want. Omni gets the real photos and the real recording.

Advice is checked against OSHA, CCOHS, WHO, NICE, NHS. No diagnosis. No made-up millimetres. Save stays on the phone.

How we built it

Expo on iPhone: camera, motion level, mic, haptics, SQLite
Hono API: Omni key stays on the server
MoveNet → packages/metrics: every displayed angle. Omni explains; it doesn’t invent
Policy layer: strips diagnoses and fake numbers before they hit the screen
Omni adapter: Yibu wants a WAV data URI; other hosts want raw base64. We verified live
Shared Zod contracts so app, API, and metrics can’t drift

Challenges we ran into

Audio encoding. Same “omni” shape, different payloads, useless errors until a live call.

Auto-capture stalls. So Capture now is always there, and the figure does not tell front from back.

Accomplishments that we're proud of

Full loop on a real iPhone: level, four views, measured recap, optional Omni voice. Photo-only and spoken questions use the same path. Sources on screen. VoiceOver, captions, 44-pt targets, local vs cloud consent. Model coaches.

What we learned

Wiring Omni for real. The key stays on the server; the phone never sees it. Sending actual photos and actual voice in one turn is different from “paste a transcript into a chatbot.” Audio encoding is per-host.

Computer vision on the stills, not a live skeleton. MoveNet reads the four JPEGs; our metrics package turns landmarks into angles. Omni can talk about those numbers.

Expo is how the app feels like a product: camera, level, mic, type, and a calm recap instead of a hackathon dashboard.

What's next for Align

Live video. Real-time feedback on basic patterns ( squat, hinge, overhead press, etc) so you’re not only judged standing still. Form cues while you move, measured from the camera. Mobility, flexibility, for coaches and trainers.