01 · CORALYST
Coralyst
On-device bioacoustic monitoring for coral reef conservation
The situation
Biorock Indonesia is a coral preservation community in Bali. Their reefs get hit by blast fishing — homemade explosives thrown into the water that kill everything in a 10–30 meter radius and can leave a reef as rubble for decades. It's illegal almost everywhere it happens, but oceans are vast, patrols are thin, and most blasts go unwitnessed. The community didn't need another dashboard — they needed a way to actually catch this happening, at sites with no reliable connectivity.
Thought process
A reef makes a sound — fish, shrimp, ambient reef noise — and so does a bomb going off. That gave us a signal to build on: a hydrophone doesn't need to be in the right place at the right moment the way a patrol boat does, and a timestamped recording is a lot harder to dismiss than a secondhand report. The hard constraint was that this had to work at the edge, where connectivity can't be assumed. That ruled out anything that depended on streaming audio to a server for inference — the detection had to happen on the device itself.
Engineering decisions
I built the AI/ML pipeline end-to-end: collecting and preparing hydrophone recordings, training a model to pick out blast-fishing signatures against the ambient noise floor of a healthy vs. degraded reef, and — the part that mattered most — getting that model small and efficient enough to run locally, on-device, with no dependency on a live connection. The system logs each detection with a timestamp and estimated location, and separately tracks reef health cues (a healthy reef crackles with life; a damaged one goes quiet) as a side signal.
My contribution
The AI/ML pipeline was mine, start to finish — model training, and the on-device deployment work to make sure something meant for a remote Balinese reef didn't quietly stop working the moment it lost signal.