Compute that
operates independently.

Digital systems often degrade in ways standard monitoring does not reliably capture. A GPU drifting before errors become visible. A network whose structure is weakening long before the event is obvious. A robotic gripper that knows an object is slipping only after it has fallen.

The common problem is that conventional pipelines sit too close to the signal they are trying to interpret. When the dynamics are subtle, thresholding is late. When the structure is rich, simple alerts miss it entirely.

The TCF programme addresses this differently. It uses a proprietary external compute layer designed to learn normal system behaviour and surface meaningful deviation early, without depending on labelled failure examples. The same core signal logic also generalises to recognition tasks: identifying events, contacts, and transitions from raw sensor streams.

The underlying signal logic has been validated in software across three independent domains. The current research programme is focused on translating that validated behaviour into low-power hardware suitable for edge, satellite, and body-worn deployment — including on-robot sensing.

Robots that handle objects need to know what is happening at the fingertip in real time. A slip starts as a small, distributed pattern across a tactile sensor; by the time conventional pipelines react, the object is already moving.

We validated the architecture on a multi-pillar tactile array spanning five objects with very different mechanical signatures — chips, paper, plastic, fabric, and a metal can. Supervised slip detection reaches near-perfect accuracy across every object. More importantly, the same architecture supports one-shot recognition: store a single slip event from one object, and the system recognises slip events on objects it has never seen, without retraining.

Cross-object transfer reaches 0.895 AUROC on average and 1.000 AUROC on the strongest transfer pairs. Inference cost is roughly 0.1 seconds per object across twelve trials in software, with a hardware path that maps cleanly to per-channel analog filter banks.

This makes the same compute layer that monitors GPUs and networks also a viable on-robot sensing primitive: low power, no retraining between objects, and a representation that generalises rather than overfits to the specific sensor calibration it was trained on.

The software validation across all three domains is complete. The coefficient package is specified for hardware implementation work. The current research programme is focused on hardware characterisation and practical deployment constraints.

The programme is structured around binary-outcome experiments. Each experiment either confirms a material's suitability for a specific computational role, or documents why it does not — both outcomes advance the design process.

We have submitted a funding application to a UK research programme and are in discussions with academic partners for formal computational characterisation. We are not yet in a position to name either.

If you are working in analog hardware, thin-film materials characterisation, physical computing, or robotic perception and are interested in the research direction, we welcome the conversation.

The TCF architecture and its applications are protected by UK patent applications covering the compute framework, monitoring approach, robotic sensing applications, and deployment-specific implementations.

The software framework is deployable today on digital hardware. The hardware programme is aimed at low-power embedded inference for environments where power, latency, or deployment constraints limit conventional digital pipelines — including on-robot perception.