HealthOS

I build computer vision systems that help radiologists catch disease. That's the day job — Applied ML Scientist. The thread running through my whole career is less glamorous than it sounds: getting models to run in places that have no business running a model.

At Toyota I squeezed a Small Language Model down until it fit inside a car's dashboard head unit. Tight memory budget, no cloud to fall back on, real-time inference. If the car has no signal in a tunnel, the feature still has to work. So it had to work offline, every time, full stop.

Before that, at Bridgestone, I put CNN models into tire plants to catch defects on the line. Cameras watching every tire come off the belt, the model flagging the bad ones before they ever shipped.

That background is the whole reason HealthOS looks the way it does. When I started it, the one thing I refused to do was send people's audio to a server. I already knew on-device was possible because I'd done it under far worse constraints than an iPhone.

It didn't start as a voice app

It started with my back. I have chronic back pain, and I wanted to find the patterns behind it — how activity, sleep, stress, and whatever else lined up with the bad days. So I built a daily tracker.

Then I walked straight into the regulatory wall. The second you correlate activity data with a health outcome, you're a medical device. FDA review. Years of validation. For a bootstrapped solo founder, that's not a hard road, it's a closed one.

So I went looking for a different signal. My criteria:

• Something your body produces on its own, without you thinking about it

• Something that actually says a lot about your physiological state

• Something that lives in wellness, not medical-device territory — this one mattered most

Voice fit all three.

Why voice

Your voice moves with your nervous system. When you're wired, exhausted, or mentally fried, it shows up in pitch, in pauses, in strain, in how fast you talk. Research labs have studied this for years — companies like Canary Speech and Kintsugi built on it. What nobody had done was put it in a normal consumer app on the phone in your pocket.

And here's the part that made it buildable for one person: voice biomarkers for general wellness aren't a medical device. Same category as an Oura readiness score or Apple Watch mindful minutes. That's the difference between shipping this year and validating it for five.

That gap is HealthOS.

What it actually does

• You talk to it for a few seconds — nothing scripted, just speak normally

• It pulls the acoustic features apart right there on the device

• It hands back nine signals: energy, fatigue, stress, mood, cognitive load, vocal strain, confidence, social engagement, and future orientation

• It tracks how those move across your days and weeks

The audio never leaves your phone. No uploads, no server, nothing to leak. That's not a marketing line — it's the Toyota lesson applied directly. You can run real models on-device if you're willing to engineer for the constraints instead of around them.

Where I am right now

• TestFlight beta is live and I'm letting people in

• First prize at the Health Wildcatters TXHCC Hackathon — they run a healthcare accelerator here in Dallas that I've just applied to

• Community — talking to other health builders on Product Hunt, cold-emailing voice researchers at UT Dallas, slowly getting to know people in the quantified-self world

Revenue so far: zero. One person, no funding, building in the open.

Try it

• TestFlight beta: https://testflight.apple.com/join/JBG3ANFF

• Discord community: https://discord.gg/SyZPw3cgG

• Website: https://healthos.live

If you're building on-device health AI yourself, I'd rather compare notes than pretend I have it figured out — so reach out either way.