Inspiration
ALS is rare, so family doctors seldom think of it first. Our case file, Alex, "probably met" the diagnostic criteria a year before his diagnosis, even though the signs were measurable: right grip 14 kg vs left 32 kg, neurofilament light chain 88 pg/mL (healthy is under 15), and lung capacity dropping from 78% to 64% lying flat. Nobody lined those numbers up. We built NeuroMotion to do it, and to make the daily measuring feel like training, not like being a patient.
What it does
A companion app takes quick daily "reps" with the phone camera and mic (hand and foot movement, speech rate). A doctor's report compares every marker against control and ALS reference ranges, takes uploaded lab results, and downloads as an APA-style PDF. It is a comparison tool, not a diagnosis.
How we built it
- Tiger Data: readings are stored in a TimescaleDB hypertable with compression and a weekly continuous aggregate. On 773,000 synthetic readings the weekly roll-up answered the same question about 3x faster than raw rows (407 ms vs 134 ms).
- Vultr: the app runs on a Vultr Cloud Compute server with automatic HTTPS at neuromotion.health.
- Gemini API: reads photos and PDFs of lab sheets into values, and writes the plain-language insight summary.
- ElevenLabs: a voice agent lets the doctor speak results into a confirm card before anything saves.
Challenges we ran into
Noisy phone measurements, real microphones, a retired Gemini model, and ambiguous reference data, which we flag in the report instead of guessing.
Accomplishments that we're proud of
A full loop from a phone rep to a database to a clinician's report, live on real infrastructure.
What we learned
Time-series data needs a time-series database at scale, and a measurement is only useful if its limits are visible.
What's next for Neuromotion
Validate against real devices, train and test a prediction model on public ALS data, add the walking test, and cover privacy and regulation before real patients use it.
