Hachi

Inspiration

Our friend George got sick just before the hackathon we had all been looking forward to. Instead of preparing to build something together, he was worrying about whether he would feel well enough to come.

He’s fine now, thankfully. But the timing stayed with us.

We already collect so much information through our watches. Heart rate, sleep, activity. We wanted that history to help us look ahead. Could it give us a little more time to prepare when our bodies start moving away from their usual patterns?

That question became Hachi.

What it does

Hachi takes your health data from devices like your Apple Watch and uses it to predict how your health might change in the future.

Instead of only showing your past and current health stats, Hachi shows predictions for specific days and times, including if you are likely to be sick, what kind of sickness, your heart rate and other health measurements. It also highlights unusual changes that could be worth paying attention to.

How we built it

We connected a Python modeling pipeline, a FastAPI backend, and a React and TypeScript interface, with forecasting inference running on the ASUS Ascent GX10.

We reviewed wearable-health research and worked with public research datasets. We compared simple forecasting methods with Chronos-2, trained prediction heads, and explored fine-tuning and wearable representations.

To make the comparisons meaningful, models received historical observations and predicted later measurements. The interface connects those outputs to the data behind them, including gaps in coverage.

Challenges we ran into

Collecting measurements and knowing what they mean turned out to be very different challenges.

Heart-rate data was available, but reliable records of when symptoms began were harder to obtain. Missing measurements and differences between devices complicated the work further.

We also had to confront a tempting assumption. A more elaborate model does not automatically make a better forecast. Some experiments failed to improve on simpler approaches, and those results helped us decide what belonged in Hachi.

Accomplishments that we're proud of

We turned concern for a friend into a working health prediction model prototype, connecting wearable data, model inference, and an interface people can explore.

We also built a way to compare predictions against what actually happened. That gave us something more useful than an impressive-looking graph. It gave us a foundation we can evaluate and improve.

What we learned

Forecasting tomorrow’s heart rate and predicting tomorrow’s illness are different problems. The second requires evidence that the first alone cannot provide.

We learned to let the results shape the product. Hachi should help people understand what their data supports, including when there is not enough information to draw a conclusion.

What's next for Hachi

Next, we want to evaluate forecasts on untouched data, strengthen support for real wearable exports, and investigate earlier illness warning using better symptom records.

The goal remains personal. We want people like George to have more time to adjust their plans, ask for help, or simply prepare for a difficult day.

A little more notice could mean a lot.