Recursive AIAt the edge
AI that builds AI for the edge.
Recursive AI means AI that builds, tests, and improves the next generation of AI models. RED KESTREL applies it to compact models for low-power devices that run without the cloud, for commercial operations and government missions alike.
Project Red KestrelBuilt faster. Built to last.
The problem
Too slow. Too big. Too fragile.
Drones, robots, sensors, and industrial equipment increasingly need AI on board. The way those models get built has not kept up.
Conventional edge AI
Time to a model
Months of hand-built work.
By the time one ships, the hardware or the mission has moved on.
Size and power
Sized for data centers.
Too large for the memory and power budget of a device running all day on a few watts.
In the field
Works in the lab. Fails in the field.
New interference, weather, data, or a new kind of attack breaks it.
Red KestrelAutomated
Time to a model
Automated, in a fraction of the time.
Candidates are generated and refined by the platform, not one engineer at a time.
Size and power
Sized for the device.
Shaped to the target’s limits from the start. Under 10 MB by design.
In the field
Stress-tested before it ships.
Proven against harsh and changing conditions, with tests that get harder as models improve.
Design targets
Small enough to live on the device.
Offline capable: models run with no cloud and no constant connection, and the data stays on the device.
- Model size
- <10 MB
- Power draw
- <10 W
- Cloud dependencies at runtime
- 0
Design targets: what the platform is built to deliver per model. Not measured results.
Platform
We don’t hand you a model. We hand you a pipeline.
RED KESTREL replaces slow, one-off engineering with an automated loop. Models are designed, shaped to the device, stress-tested, and verified before they ship. Three stages run in a secure central environment; the fourth lands on the device.
Value
Five things you actually get.
The same outcomes for commercial operators and government programs.
New or updated models in a fraction of the time.
Automated design and testing replace months of manual development, so a model can keep pace with the hardware, the data, and the mission.
The REY way
Integrated beats assembled.
Others
- Separate vendors for design, optimization, and testing.
- One-off models that start aging the day they ship.
- Cycles measured in months.
- Tested in the lab. Surprised in the field.
The REY way
- One pipeline from design to deployment.
- Continuous improvement: every test feeds the next model.
- Hardware-aware from the first candidate.
- Verified by automated checks and human review before it ships.
Industries & missions
One platform. Two markets.
The same compact, tested models serve commercial operations and government missions. Dual use is the design, not an afterthought.
0108
Hardened for the toughest missions. Priced and scaled for commercial use.
Also: critical infrastructure networks, and medical and consumer devices that must keep personal data on the device.
Prior work · Recorded replay
Proof of execution.
This recorded replay of earlier work in network anomaly detection shows how we build and validate AI: error targets set in advance, testing on data the system never saw, safety rules the AI cannot override, and honest reporting of false alarms. The same discipline drives how RED KESTREL verifies every model before release. It applies equally to enterprise and government networks.
- Error targets set in advance, then measured against
- Tested on a day of traffic the system never saw
- Safety rules the AI cannot override
- False alarms counted and reported, not hidden
What this replay shows
The recording is a replay of one day of real network traffic, sped up, run through an anomaly-detection model in detect-only mode. The system recommends; it never acts.
- Top row: performance measured on a held-out day the model never trained on, including the false-alarm rate against the target set before evaluation.
- Middle: running counters for alerts blocked by the safety rule, confirmed detections, coincidental flags, and false alarms, above a risk-score timeline against the alert threshold.
- Bottom left: the current assessment for one server, with its risk score, threshold, and the recommended action for an analyst.
- Bottom right: the event log, every alert newest first, nothing hidden or truncated.
On small screens the text in the footage is hard to read; this description covers the same content.
Replay results · recorded traffic
25
Confirmed detections
Real attack traffic correctly identified.
21
Blocked by safety rule
Actions against protected servers, stopped before execution.
17
False alarms
Normal traffic flagged. Our measured error rate.
Roadmap
Where it stands.
RED KESTREL is our flagship research and development platform. This is its status in plain terms.
Phase I
Automated design and optimization core
The central environment that generates candidate models and shapes them to a target device’s limits.
In development
Phase II
Adversarial testing suite
Automated stress-testing against interference, noise, weather, changing data, and deliberate attacks, with tests that harden as models improve.
Planned
Phase III
Field validation with pilot partners
Verified models on real hardware with commercial and government pilot partners, with results reported transparently.
Planned
Work with us
Two ways in. One platform.
Government programs and commercial operators engage differently. The platform, the engineering discipline, and the team are the same.
Government
Direct, or through your prime.
We work with agencies directly and as an AI/ML teaming partner to prime contractors.
Verified on fielded hardware.
Models proven against degraded conditions and reviewed by a person before release.
Built for programs that cannot assume the cloud.
Defense, federal civilian, and state and local missions where connectivity is limited or not permitted.
UEI, CAGE, and capability statement available on request.
Commercial
Start on the hardware you own.
A scoped pilot on your equipment and your data, with success criteria agreed up front.
No cloud bill. No connectivity requirement.
Models run on the device, so the economics hold at the edge.
Models that keep improving.
Updates come from the pipeline, not from a new engineering project each time.
Registration and classification
- SAM.gov status
- Active
- Ownership
- Woman-owned · Minority-owned · Small business
- NAICS
- 541511 · 541512 · 541715
- 541511 Custom Computer Programming Services
- 541512 Computer Systems Design Services
- 541715 R&D in Physical, Engineering, and Life Sciences (Nanotechnology / AI R&D focus)
Contracting pathways — We engage through SBIR/STTR, Other Transaction (OT) agreements, commercial solutions openings, and subcontracts with prime contractors.
Commercial engagement models
Pilot programs
A scoped pilot on your hardware and your data, with success criteria agreed before work starts and a verified model at the end.
Custom model development
Compact models built for a specific device and task, delivered as ready-to-run software with the test results that back them.
Platform licensing
For teams that want to run RED KESTREL in their own environment, as the platform matures.
Joint R&D
Shared research on new device classes, sensors, or operating conditions, with results both sides can use.
Built for the edge. Proven before it ships.
Tell us about your hardware and your conditions. We will tell you what a verified, on-device model looks like for it.
Request a BriefingContact
Start the conversation.
Direct Inquiries
For briefings, pilot discussions, capability statements, or teaming on a proposal, use the form alongside, or email directly for a faster route. Commercial and government inquiries are handled by the same team.