Aligning AI to the human brain.
We turn involuntary biosignals into per-token reward signals for the next generation of AI.
Our vision
The signal AI has been missing.
Frontier models are trained on the shadows of human judgment: clicks, thumbs, star ratings, written rationales. The signal that actually matters is the source: what the human brain detects, evaluates, and rejects. These responses are involuntary, information-dense, and personalizable.
Reinforcement Learning from Biofeedback (RLBF) captures them. We pair EEG with other biosignal modalities to produce continuous, per-token, per-frame, per-second reward signals that no annotation pipeline can replicate. It's faster than RLHF, denser than preference labels, and reveals subconscious effects like the uncanny valley.
What we are building
Model to capture stack to alignment platform.
Model
001A foundation model that maps multimodal biosignals such as EEG to per-token error and preference signals across text, audio, and video.
Capture Stack
002Lab-grade and consumer-grade hardware that records biosignals time-locked to every token a human reads, with millisecond-accurate stimulus alignment.
Alignment Platform
003The pipeline from raw biosignal to a JSONL reward stream plugged directly into the RLHF stacks of the labs training the world's most capable models.
