Our engineers' lab work, written up in full: the set-up, the data, and what it does and does not show yet.
Where does VAHA turn speech into text, and with which model? The device transcribes its own captures with faster-whisper base.en; notes recorded on the phone can reach a cloud service; and in a first 120-word test, base.en and small were one word apart, with small about 2.9 times slower.
Can a bad sensor reading pass as a good one? How VAHA's temperature, humidity, TVOC and water-flow sensors are wired to an Arduino UNO Q and reach the app, why a failed temperature or humidity read is caught, and the two ways TVOC failures and damaged replies are stored as readings.
Does every voice note reach the phone intact, and only the phone? How VAHA’s app pulls captures from the device over the local network; three gaps in the current build, from no checksum comparison before the device copy is deleted to plain HTTP; and a checklist for local-only sync.
How reliably does VAHA catch its wake word, “Marvin”, and its stop phrase over a bathroom fan and a running tap? What the keyword-spotting model was trained on, its 97.6 % validation accuracy, and why that is not yet a figure for the device itself.
Do four low-cost water sensors give usable signals before any calibration? First bench results for turbidity, TDS, pH and level modules on an Arduino Uno R4 WiFi, and the six rules they set for the calibration that comes next.
Can a hub and its sensor nodes prove who they are to each other over Bluetooth LE without pairing, MAC addresses or shared secrets? Lab results from a certificate handshake on two ESP32-C3 nodes, including a 41-hour unattended run.
No whitepaper matches these filters yet.
