Ninety-Nine%

Inspiration :

A Parkinson's patient sees a neurologist for about 20 minutes, every 3 to 6 months. Symptoms change every hour with the levodopa cycle. Doctors adjust doses based on what the patient remembers, and patients with speech or memory decline remember poorly. We built Ninety-Nine% to cover the 99% of the patient's life the doctor never sees.

What it does : Ninety-Nine% turns a wrist band into a continuous neurology follow-up.

The band records motion, voice and light all day.
We detect tremor, dyskinesia, posture, walking and speech slowing.
We match each event to the medication schedule.
The doctor gets a one-page synthesis with concrete changes, for example "move the 17:00 dose to 18:00".
The patient gets a reminder on the wrist. At the next dose, the system checks that the change worked.

How we built it :

Next.js and TypeScript dashboard, with a three.js 3D replay of real tremor episodes.
Python bridge that streams band data over Bluetooth.
On-band signal algorithms with false-positive rejection, covered by 41 automated tests.
Deepgram for speech rate, Gemini for the clinical synthesis (with an offline fallback), ElevenLabs for voice cues.

Challenges we ran into :

False positives. Brushing teeth looks like tremor. Swinging an arm looks like a double-tap. We logged raw sensor data and replayed it offline until the detectors ignored normal movement.
Wide tremor range. Real tremors run from slow, wide shakes to fast 12 Hz ones. One detector had to catch both.
Hardware limits. The band has no vibration motor, speaker or heart-rate sensor. We used LED signals and laptop audio, and we label every simulated metric as simulated.

Accomplishments that we're proud of :

The full loop works on a real band: capture, detect, recommend, act, verify.
A doctor can replay a tremor episode in 3D from the patient's own wrist motion.
Voice becomes a measurable signal: we track speech rate, pauses and volume over time and alert under 100 words/min.
Every metric shows whether it is real or simulated.

What we learned :

Doctors don't want more data. They want one decision tied to the dose schedule.
Most of the work in wearables is rejecting false signals, not detecting true ones.
Speech is an underused biomarker, and anyone with a microphone can collect it.

What's next for Ninety-Nine% :

About 10M people live with Parkinson's and 55M with dementia, and most are followed with a few short visits a year.
Next 3 months: pilot with 2–3 neurology clinics and measure how many dose changes our recommendations drive.
Validate the tremor and speech detectors against clinical scores (MDS-UPDRS).
Move to hardware with a vibration motor and heart-rate sensor.
Start the medical device regulatory path. Ninety-Nine% is not a medical device today.