Streaming EEG at low latency

EEG is microvolts riding on noise. The job of an acquisition pipeline is to get those microvolts off the head, say which parts are real, and land them somewhere you can trust months later — fast enough to be useful live.

Do only the reversible work at the edge

On-device processing is limited to things you’d never regret: DC removal, a configurable notch for mains, and band-limiting. Anything heavier — the calls you might want to redo — stays off-device where it can be re-run against the raw stream. The truth survives.

Let the flags travel with the data

Each window carries cheap artifact flags: saturation, motion, impedance. Because the quality call is attached at capture time, every downstream consumer makes the same decision and nobody re-implements quality detection three different ways. The full-rate stream is written losslessly with metadata, so a session is reproducible long after it ends.

Filter reversibly, decide later. Record quality as flags, not as bakes.

It’s an unglamorous pipeline on purpose. Under forty milliseconds glass-to-stream, no gaps, and a recording you can still believe when you open it next quarter. That last property is the one people forget to design for.