Flynn is an embedded anomaly detector that turns equipment telemetry into usable local intelligence. It runs on bare-metal in 8,480 bytes, self-calibrates from normal behavior, and delivers deterministic, auditable detection for machines, tools, and edge systems where cloud-first AI latency is too high, too opaque, or unavailable.
From sub-sea to deep-space, equipment talks in telemetry. Vibration, current, temperature, pressure, flow, timing. The signals are there. The local, deterministic, auditable early detection rarely is — a detector small enough to embed, smart enough to self-calibrate, and reliable enough to act on without a cloud stack, a labeled dataset, or an ML team behind it.
Telemetry exists. A detector you can trust close to the source does not. The decision lives far from the action it informs.
Supervised approaches demand labeled failure data. Most equipment is designed not to fail. The training set does not exist.
Most consequential equipment lives where the cloud doesn't. A cloud-dependent detector is a detector that doesn't work.
Tools built for dashboards do not belong in firmware. Too large, too dynamic, too dependent on external services.
Designed from first principles for MCU-class hardware, for enterprise-scale deployments. The numbers below are entry conditions.
Not 8 GB or 8 MB — eight thousand four hundred and eighty bytes. Fits on the MCU already inside the equipment.
Self-calibrates from normal behavior in under two seconds at kilohertz-class rates. No labels. No tuning.
Zero dynamic memory allocation after init. Hard requirement for safety-critical firmware — achieved by design.
The entire detector is a single human-readable C source file. No model blob, no binary, no opaque runtime.
Zero false positives across 120 hours of bearing-class soak. Bounded above 0.025/hr at 95% confidence.
Deterministic across runs and silicon. Replayable. Auditable. Eligible for safety-certified voting architectures.
The signal dominates at the moment of conflict. The prior persists. The two never share the wire.
Install. Deploy. Walk away. After 1,700 samples, the threshold is locked — and an alert from Flynn means a real and sustained departure from the conditions present at installation.
For the first 1,700 samples after deployment, Flynn observes the signal without producing alerts. It accumulates a statistical model of the signal's normal behavior — variability, autocorrelation, distribution shape, dynamical character — and commits a conservative detection threshold.
Every subsequent sample passes through the detection pipeline and is scored against the calibrated model. The locked threshold refuses adaptive drift by construction: a slow-developing fault cannot be quietly absorbed into the new normal.
Five-domain validation on the same binary and works across signal domains that share no domain assumptions.
Each stage is buildable and useful on its own. Each is composable with what comes next. The structural-separation property propagates through every layer.
Deployable today as embedded firmware. Validated across five domains. Ready for production integration into industrial equipment.
Three to ten Flynn detectors across one asset, with a fabric node maintaining operational-regime memory. Detector layer remains authoritative.
Reflex-class response. Cerebellar memory of operational regimes. Cortical decisions across the site. Forensic replay back to any moment.
From a single sensor on an engineer's bench to a thousand-asset fleet in an air-gapped facility. Specific terms are formalized through direct conversation.
For developers, researchers, and innovation groups testing Flynn against their own data.
For manufacturers embedding Flynn into commercial equipment at production scale.
For operators deploying Flynn across fleets, retrofit or new-build, on or off network.
Evaluation access, OEM licensing, enterprise deployment — every engagement starts with a direct conversation. We don't sell you a thing and walk away. We sell you a thing that keeps working when you need it to, for as long as you need it to.
