Field instrument for hearing an outbreak before it bites.
ON‑DEVICE · OFFLINE
no audio leaves
this phone
01
The model
~0.3 MB · on‑device
Pick model.json, the .bin shard, and model_meta.json from your Phase 0 export
No model loaded.
02
Bench test
skip the microphone
Drop in a .wav and classify it directly. No speakers, no room, no mic.
Load a model first. Then feed it a clip from your Wingbeats test set.
03
What the model hears
live spectrogram
0–2000 Hz
1000
300
Mosquitoes live between the two guide lines. Flat and empty means the mic is not picking anything up.
Microphone off.
04
The reading
—Waiting for audio
Confidence gate0.60
Your Phase 0 sweep: 0.60 keeps 74.8% of clips at 83.0% accuracy. Raising it to 0.90 buys 93.7%
but throws away 61% of what you hear.
05
Where it is active
Location offNetStorage
Every confident detection is stamped with a location and time and written to this device.
Nothing is uploaded. Turn off the network and it keeps working.
Total0
Aedes0
Last 24h0
06
Ward risk
detections × water × rain
Rain: not set
Rain slider stands in for a weather feed (mm, last 7 days). Wire it to a real API later; for a demo,
set it by hand. Breeding needs standing water, so rain multiplies the risk from reported sites.
No wards scored yet. Log a detection or report a water site, and a ward appears here with a task for a health worker.
07
Detection log
Time
Species
Location
Conf
Nothing logged yet. Confident detections are saved here and survive a reload.