Equipment talks. Flynn listens.
Whitepaper · Q2 2026 · 11 sections
A whitepaper on deterministic, embedded anomaly detection — in telemetry, at the edge of the world.
Contents
- Executive summary
- The telemetry era
- The four gaps
- Why Flynn is different
- How Flynn works
- Architecture of trust
- Validated across domains
- Where Flynn runs
- The Flynn roadmap
- Tiers and engagement
- About Flynn
Executive summary
Industrial equipment fails for reasons that show up in the signal long before they show up in the world. A bearing wears. A pump cavitates. A motor drifts out of phase. The data is there. The listener is not.
This whitepaper introduces Flynn, an embedded anomaly detector that turns equipment telemetry into local intelligence. Flynn runs bare-metal on the microcontroller already inside the machine, in 8,480 bytes of deterministic C. It calibrates itself from normal operation in 1,700 samples † — under two seconds at kilohertz sampling rates — with no labels, no cloud, and no per-deployment tuning.
Flynn monitors equipment telemetry at the source.
The telemetry era
The signal is everywhere
From sub-sea to deep space, telemetry talks. Vibration. Current. Temperature. Pressure. Flow. Timing. Every rotating asset, every fluid system, every electrical bus, every thermal envelope emits signal continuously. A modern industrial site instruments ten thousand sensor channels, each emitting hundreds to thousands of samples per second.
The listener is missing
The infrastructure that turns telemetry into action is overwhelmingly centralized: a sensor reads, transmits, and accumulates in a historian or cloud data lake, to be analyzed later by people or models far from where the signal began. A large and growing class of applications cannot accept this, for structural reasons:
- Connectivity is unreliable. The most consequential equipment runs where backhaul is intermittent, slow, expensive, or absent.
- Decision latency matters. A dry-running pump destroys itself in seconds; the decision cannot wait for a round trip to the cloud.
- Trust cannot be outsourced. Decisions affecting safety, compliance, or product integrity must be auditable and reproducible.
- Cost-per-decision matters at fleet scale. Sending every sensor reading to a cloud collapses economically at ten thousand assets.
THE CLOUD MODEL
[ sensor ] --> [ gateway ] --> ~~ network ~~ --> [ model ]
|
[ operator ] <-- [ dashboard ] <-- ~~ network ~~ <---+
round trip: minutes to hours
breaks wherever connectivity, latency, power, or
radiation break the chain
FLYNN
+--------------------------------------------------+
| THE EQUIPMENT |
| |
| [ sensor ] --> [ MCU :: Flynn, 8,480 B ] --+ |
| | |
| [ local alert ] <--+ |
+--------------------------------------------------+
round trip: none
microseconds, offline, deterministic, hardware under $10
The four gaps
Conventional anomaly detection leaves four gaps. Each one keeps a class of equipment uninstrumented today.
High decision latency
The signal travels to a distant server and the verdict travels back. For a pump that can wreck itself in seconds, the round trip is the failure.
No labeled training data
Supervised models demand labeled failures. Industrial equipment runs for years between them, and when it fails the event is rare, costly, and often catastrophic. The labels never accumulate.
Equipment lacks connectivity
Cloud-first systems assume connectivity; the most consequential equipment runs without it. The cloud is slow, or gone exactly when it is needed.
Monitoring tools exceed firmware constraints
Detection tools built for dashboards are too large for firmware. In embedded systems, determinism and a small footprint are entry conditions.
Flynn closes that gap in 8,480 bytes — self-calibrating, embedded, bare-metal anomaly detection that turns telemetry into deterministic, auditable intelligence.
Why Flynn is different
8,480 bytes
Flynn’s runtime footprint is 8,480 bytes — eight thousand four hundred and eighty, not eight kilobytes. The figure decides the silicon it runs on, the equipment it embeds inside, and the cost per node across a fleet.
Self-calibrating
Flynn learns the statistical signature of normal operation from the signal’s first sampling period — the only input it requires. The integrator’s workflow: install, power on, walk away.
Deterministic
Flynn produces bit-identical output for bit-identical input, across runs and across silicon. This determinism is the structural property that allows Flynn to participate in safety-certified voting architectures, to be replayed forensically after an incident, and to be verified against published test vectors.
Auditable
Flynn is auditable by construction. Because it is deterministic, its behavior is fully reproducible and independently verifiable against published test vectors, and it ships with certification-ready artifacts for safety review. For certifying authorities, the full source is available under NDA or in escrow.
Bare-metal
Flynn runs on bare metal. Its only dependency is a basic math library — no operating system, no network stack, no filesystem, no scheduler.
How Flynn works
Flynn’s operating model is two phases. Learning. Detection. The transition between them is automatic and one-way.
peak amp
|
|
| #### <-- ALERT
| ####
| #### ####
|- - - - - - - - - - - - - - - - - - -#### -#### - <-- locked band
| #### #### ####
| #### #### #### ####
| #### #### #### #### #### #### #### ####
| #### #### #### #### #### #### #### ####
| #### #### #### #### #### #### #### ####
+------------------------------------------------
s1 s2 s3 s4 s5 s6 s7 s8
|<--- ENROLL --->|<--------- DETECT ---------->|
LOCK
Learning
For its first 1,700 samples †, Flynn watches in silence and builds a statistical model of the signal’s normal behavior. At the end of enrollment it commits one detection threshold and locks it. The threshold does not move, adapt, or drift.
Detection
After enrollment, every sample passes through the detection pipeline. A score below the threshold is normal; a score at or above it is anomalous. The magnitude of the score measures the severity of the deviation.
The locked threshold is a deliberate design choice. Adaptive thresholds carry a well-documented failure mode: a slow-developing fault gradually drifts the baseline upward, the threshold follows, and the fault dissolves into a new “normal.” Flynn refuses that by construction.
† Samples required for complete enrollment will vary by equipment type, state, and operational envelope.
The architecture of trust
Evidence outranks the prior
Every learning system must hold two things at once: the accumulated experience that sharpens its judgment, and the direct measurement that reveals when something has changed. The common failure is that experience overrides the measurement.
Flynn separates them by locking
Flynn’s locked-threshold design enforces that separation. After deployment, no feedback path runs from accumulated experience back into the classifier; the judge of each new sample stays isolated from any later influence.
Once the equipment’s nominal signal envelope is learned through enrollment, it’s locked in permanently. When the equipment telemetry trends towards a threshold it does not change the enrollment data.
Validated across domains
Flynn has been validated across five distinct industrial signal domains, with zero configuration changes between them.
- Bearing vibration — CWRU benchmark
- Precision 1.0 · Recall 1.0 · F1 1.0
27 fault pairs (inner race, ball, outer race × 3 severities), default operating point. - Run-to-failure — NASA IMS bearing dataset
- Lead time ~17 days · failing bearings 2 / 2
30+ days, 4 bearings, 8 channels. Adjacent healthy-bearing channels registered the same developing fault at proportionally lower magnitude — cross-channel sensitivity with correct severity ranking, not a false positive. - Ambient & diurnal
- 0.074–0.080 false positives per hour
336 hours of environmental and process-control signals. Fewer than two alerts per day — compatible with normal shift-cycle review. - Electrical grid stability
- F1 0.532
10,000 instances, 12 features. Same source, same flags, no per-domain configuration. - Soak — synthetic vibration
- 0 false positives
120 hours, 5 seeds × 24 hours. 95% CI ceiling < 0.025/h.
Zero false positives on bearing data. Operationally low on diurnal. Full protocols and metrics.
Where Flynn runs
Any compliant microcontroller
Flynn is delivered as a compact binary, compiled for your target from a single C99 codebase. It builds with any C99-compliant toolchain. It has been hardware-validated on commodity 32-bit MCU architectures including both proprietary and open instruction sets.
Bare-metal C
Flynn runs on bare metal: no operating system, no network stack, no filesystem, no scheduler, and no memory allocation after initialization.
Zero heap
Flynn fixes its entire working set at compile time and allocates nothing afterward — the hard requirement of safety-critical firmware, where certification forbids dynamic allocation.
From sub-sea to deep space
The same detector compiles to commodity industrial silicon, radiation-tolerant microcontrollers, and power-constrained sensor nodes.
[ ORBIT ] [ DRONE ] [ VALVE ] [ SUBSEA ] [ IMPLANT ]
| | | | |
+----------+----------+-----------+-----------+
|
( F L Y N N :: 8,480 bytes )
one binary
no domain-specific model
no per-domain retraining
The Flynn roadmap
STAGE 01 :: THE NEURON today, shipping
[ signal ] --> ( Flynn ) --> [ score ]
one detector, one signal, one decision
STAGE 02 :: THE DETECTOR FABRIC 2026, pilots engaging
( Flynn ) --+
( Flynn ) --+--> [ coordination ] --> [ asset state ]
( Flynn ) --+ regime memory
3-10 detectors, one asset. Detectors stay authoritative.
STAGE 03 :: THE NERVOUS SYSTEM multi-year
[ CORTEX ] site-wide decision-making
^
[ CEREBELLUM ] operational-regime memory
^
[ REFLEX ] sub-millisecond, at the equipment
Today · the neuron
The Flynn detector is the foundational unit: one signal in, one score out. It ships today as embedded firmware, validated across five domains.
In development · the deterministic, auditable, neural network
A coordination layer that composes multiple Flynn detectors on one asset — three to ten, covering different sensors, with a fabric node that remembers operational regimes. Pilot deployments begin in 2026.
Beyond · the nervous system
A facility-wide distributed cognition. Reflex-class response to sub-millisecond events. Cerebellar memory of operational regimes. Cortical decision-making across the whole site. Long-term memory with forensic replay back to any moment in the equipment’s history.
Tiers and engagement
Flynn supports several engagement models, scaled to the user. Whatever you need, let’s talk.
- Evaluation access — no-cost, bounded scope, NDA required.
- OEM licensing — per-unit royalty, integration support, long-term source stability.
- Enterprise deployment — direct support, custom integration, compliance reporting, multi-year maintenance.
About Flynn
Flynn is built by EntroMorphic, an embedded-systems and machine-intelligence company founded on one conviction: the most consequential equipment in the world should stay safe without the cloud. Its engineering leadership brings deep experience in high-performance computing, embedded firmware, and applied machine-intelligence architecture.
Flynn proves a principle: small, deterministic, auditable, structurally-separated detectors serve industrial equipment better than the centralized analytics that have dominated the past decade.
Next steps
For evaluation access, OEM licensing, or enterprise deployment conversations: tripp@inlikeflynn.io