AedesCapella

An embedded acoustic mosquito classification system with ESP32 hardware and a Supabase-backed operator dashboard.

Overview

AedesCapella processes audio for the wingbeat signature of Aedes aegypti. A detection result is reported to a separate controller that manages citronella repellent delivery. The system is designed for a fixed installation.

My contributions cover embedded integration, deployment, device-side processing, testing, and investigation of inference behavior.

Two controllers

Detection

ESP32-S3

938 ms per inference, measured on device

Captures audio, resamples it, computes a mel filterbank, and runs the quantized model. It reports the result to the system.

Actuation

ESP32-C3

Exclusive relay authority, fails closed

Holds the relay connection and defaults to closed. Separating relay control from audio processing lets the actuator keep its own control logic.

Signal path Microphone to dashboard

Stage	Runs on	Note
Capture	ESP32-S3	Onboard microphone, continuous
Resample	ESP32-S3	Fixed point, parity tested against a golden vector
Mel filterbank	ESP32-S3	Precomputed filter coefficients
Inference	ESP32-S3	Quantized 2D-CNN, 938 ms per window
Relay decision	ESP32-C3	Fail closed, exclusive authority
Reporting	Supabase	Events to a React operator dashboard