One active program — Sentio, the sensing node that gives machines ears, eyes, and an on-device brain. The research below is what we're actually working on right now, stated plainly: what's validated in simulation, what isn't yet, and what we'll publish as field data lands.
The hard part of acoustic drone detection isn't the concept — it's making a small, cheap, passive node hold up outdoors against wind, distance, and look-alike sounds. That breaks into three research thrusts:
Detectors trained on physics-based simulation and synthetic mixtures, then hardened against real recordings. The open question we're honest about: how much does sim performance degrade at a real range, in real wind? That number comes from the field campaign — and we'll publish it either way.
A node you can mount anywhere means a microphone aperture measured in centimeters — which makes direction-of-arrival estimation at distance a low-SNR problem. We're studying how much bearing accuracy small geometry can actually buy, and where the physics says stop.
The runtime that points a camera at a sound, fuses both opinions into one track, and publishes it in standard formats — on commodity edge silicon, no cloud in the loop. This layer is designed to be reusable: the same loop that hears drones can hear a failing bearing or a fence-line intrusion.
We kept the discipline from our compression era, where every public number shipped with a reproduction path. For Sentio that means:
The drones that worry every air-defense planner right now emit no radio signal at all — and the counter-drone industry is largely built on hearing radios. A plain-language note on the threat, why passive acoustics is the right cue, and what we're building to prove it.
Lab notes from both eras — current Sentio work and the archived compression research.
From 2025–26 Sipsa Labs worked on near-lossless LLM weight compression. The product was discontinued in June 2026; the verified results, methodology, and packs remain public — that's what "built in the open" means when a chapter ends.