Brain foundation models
Our most valuable system, still the least understood.
Pretrained on 3 Million hours of EEG
2× GPT-3 pretraining token count, and over 1,000× the largest published EEG foundation model.
Pretraining token count
Piramidal
2×
GPT-3
1×
20B parameters
12× the largest published EEG foundation model, 8× the largest published time-series model.
Largest published EEG foundation model≈1.7B
Largest published time-series model≈2.5B
Piramidal20B
≈1.7B Largest published EEG foundation model
≈2.5B Largest published time-series model
Invariant representations
Transfers across montages, hardware, subjects, and conditions with minimal fine-tuning.
Montages
Hardware
Subjects
Conditions
Universal Neurosensor Model
Bigger models reach lower loss: each size finishes below the one before it.
Two products on one foundation model
Marcy
A clinical-grade EEG agent that monitors, flags and reviews EEG recordings in real-time.
Pending 510(k) submission, not available for sale in the U.S.
Fabryc
Enterprise AI infrastructure for real-time tracking of hundreds of EEGs.
Launching in 2027.
Frontier model scale, clinical grade performance
General intelligence
Performs across devices, subjects and indications.
Inductive transfer learning
Fine tune to any downstream task, with cross-condition (ICU ↔ epilepsy) and cross-modality (scalp EEG → intracranial) generalisation.
Label efficiency
Minimizes the requirement for human labels, decreasing the needed label count by 5,000x within the XL foundation models.
Channel-agnostic
Adapts to any number of channels, montages, hardware devices, and sampling frequencies.
3–10x patient review volume
In a pilot, using our model triples monitoring throughput, scaling the technologist-to-patient ratio from 1:4 to 1:12 (1:64 tba).
Enterprise neurology infrastructure
Designed for real-time AI at 1,000x patient scales, our infrastructure connects multi-site neural data into a single, device-agnostic platform.