I founded Cerebralx and led Xone, a lower-cost brain–computer-interface prototype that captured visual-attention EEG patterns, classified one target from a fixed set, and mapped it to a predefined connected-device command.
The product question was whether lower-cost brain-signal hardware could support another control path: a person focuses on a visual target, the system recognizes the associated neural response, and a connected device receives the intended command.
Dry electrodes, constrained electronics, movement, environmental noise, individual variation, real-time processing, and device connectivity all affect whether an apparent signal becomes a dependable command. No single model or hardware component can solve the loop in isolation.
I owned the stack across electronics and PCB prototyping, embedded firmware, signal processing, SSVEP decoding experiments, real-time control, and cloud/device integration while coordinating the multidisciplinary team, grants, demonstrations, and stakeholders.
The default story stays at the level of the person and the work they are trying to finish.
The interface presents distinguishable visual frequencies tied to possible commands.
Dry-electrode EEG hardware records the activity associated with sustained visual attention.
Filtering and signal processing reduce interference before interpretation.
SSVEP-oriented classification matches the response to one frequency from the fixed target set.
The application sends the chosen command through the embedded and IoT path.
A better classifier cannot recover a signal the electrodes never captured, and better hardware is not useful if the command path feels slow or opaque. The team treated the complete signal-to-action loop as the product rather than optimizing one layer in isolation.
EEG acquisition, noise reduction, classification, embedded control, and connected-device behavior could each look promising in isolation.
A demonstration failed whenever any boundary between those layers introduced noise, latency, an invalid command, or an integration mismatch.
I led hardware and software iteration as one system: electrode and ADC work informed filtering; decoding constraints informed the interface; firmware and connectivity completed the command path.
The durable engineering principle was multidisciplinary: when the input comes from the physical world, product reliability lives between disciplines.
Government-backed support associated with the Cerebralx venture period.
Ayush founded and led work spanning hardware, embedded systems, signal processing, software, and product.
The development work included a multiplayer BCI demonstration and connected-device control loop.
The technical route traces signal acquisition, constrained decoding, connected-device execution, and the publication boundary between a working prototype and a clinically validated device.
The system connects visual stimulus, EEG acquisition, filtering, decoding, embedded control, and device response.
Classification quality depends on electrode contact, acquisition, filtering, and the interaction protocol before inference starts.
The product must account for weak confidence, calibration variation, connectivity loss, and device-state mismatch.
Demonstration, funding, and multidisciplinary implementation do not establish clinical performance.
Have a useful AI idea trapped in an unclear brief or brittle prototype?
Send me the rough version →