No cloud in the loop
Inference completes on the device. There is no upload path to secure, no telemetry to disclose, and no service to keep paying for.
Applications
Six industries with the same problem in different clothes
Most problems are forced onto hardware that was never shaped around them. Fixed power budget. Unreliable or absent network. A decision that has to land before the next control cycle. Wherever those three meet, moving weights across a bus stops being affordable.
What the six have in common
Nothing to reload on wake
Weights are non-volatile and live in the array. A node can sleep for days and be ready on the first sensor event.
Degrades, does not die
Work is distributed across many small arrays. A damaged region costs precision in one place instead of taking the system down.
Sits beside what you have
The engine is a co-processor on your existing MCU or SoC. It takes the perception workload; your host keeps the application.
Automotive
Perception that stops when the connection stops is not perception.
Perception that keeps working when the link does not
Sensor fusion, driver monitoring and predictive diagnostics that run inside the ECU on a power budget a vehicle can actually spare, with a wake-up that costs nothing, because the weights never left the array.
- The binding constraint
- Thermal and power headroom in a sealed enclosure, plus a hard real-time control loop.
- What changes
- Inference stays inside the vehicle: no upload, no round-trip latency, no connectivity assumption.
Four streams, one decision, one enclosure
- Camera and radar front-end perception
- In-cabin driver and occupant monitoring
- Predictive maintenance from vibration and current signatures
- Always-on wake-word and anomaly detection
Aerospace & Defence
Bandwidth is narrow, latency is long, and the connection may be jammed or gone.
Autonomy for platforms that are on their own
On-board processing for airborne and orbital systems where the downlink is narrow, contested or absent, and where a single-event upset must degrade the answer, not end the mission.
- The binding constraint
- Severe power and mass budgets, radiation environment, no operator in the loop.
- What changes
- Distributed tiles degrade gracefully instead of failing hard, and nothing needs to be transmitted to be understood.
A strike, and the answer that survives it
- On-board image triage before a constrained downlink
- Guidance, navigation and terrain-relative perception
- RF and acoustic signature classification
- Health monitoring across long unattended deployments
Robotics & Physical AI
A robot that has to ask a server what it is looking at reacts too late.
A control loop that closes in the robot
Reflex-rate perception and control for manipulators, mobile robots and drones, where a decision that arrives 200 ms late is not a decision at all.
- The binding constraint
- Battery-limited duty cycle and a control loop measured in milliseconds.
- What changes
- Perception, adaptation and actuation share one low-power fabric on the robot itself.
The gap, closed every cycle
- Tactile and proprioceptive sensing at reflex rates
- Visual servoing and obstacle avoidance
- On-device adaptation to wear, payload and surface change
- Multi-agent coordination between fleet members
Healthcare & Medical Devices
A monitor that uploads a patient in order to decide has already made the decision somewhere else.
Continuous monitoring that never uploads a patient
Wearable and implantable devices that classify physiological signals on the device, where privacy is a property of the architecture rather than a clause in a policy document.
- The binding constraint
- Years of battery life, strict biocompatible thermal limits, and regulated data handling.
- What changes
- Raw physiological data never leaves the patient, which removes an entire class of compliance surface.
Classified where it is measured
- ECG, EEG and EMG classification on-device
- Continuous glucose and vitals trend detection
- Hearing and prosthetic signal processing
- Point-of-care diagnostic instruments
Industrial & Energy
Most of what a sensor sends is routine. Only the exception matters.
Condition monitoring on machines that were never networked
Retrofit intelligence for factory floors, grid assets and remote infrastructure. Nodes that run for years on a harvested or battery supply and report an event, not a stream.
- The binding constraint
- No reliable network, no maintenance access, and a decade of expected service life.
- What changes
- The node ships a conclusion over a narrow link instead of hauling raw signal to a data centre.
A sideband, where there was none
- Vibration, acoustic and thermal anomaly detection
- Motor and bearing failure prediction
- Quality inspection at the machine
- Grid and pipeline edge telemetry
Consumer & IoT
A device that senses continuously is judged not only by how fast it responds, but by how efficiently it waits.
Always-on intelligence inside the device
Wake-word, presence, gesture and context awareness that stay on continuously without the power draw or the privacy cost. Because the array is transparent, it can sit behind the display.
- The binding constraint
- Bill-of-materials pressure, tiny thermal envelope, and unforgiving battery expectations.
- What changes
- Features that previously required a cloud round-trip become local, instant and private.
Sensing through the screen in front of it
- Always-on keyword and audio scene detection
- Presence, gesture and gaze sensing
- On-device personalisation that never syncs
- Display-integrated sensing
Not on the list
Your application probably still fits the same shape
If your workload is dominated by matrix-vector products, runs continuously on a constrained supply, and cannot depend on a network round-trip, it belongs in this set. Whatever the industry is called. An evaluation will tell you within weeks, including when the answer is no.
Working at the limits of conventional computing? Discuss your workload.
Tell us what the workload has to do, and where it has to run. An evaluation starts there, with your power envelope and your deadline, not with a datasheet.