For small uncrewed aircraft

The problem isn't the decision.
It's proving it afterwards.

“The pre-flight go/no-go decision is not hard. I’ve made it ten thousand times. That decision is not my problem. My problem is proving it afterward. … A records request. A council question. A waiver renewal. And a deposition eighteen months later asking which version of the policy was in force — today, honestly, I’d have to go look.”
A chief pilot, eight years running a drone-as-first-responder programme

So we do not stop your launch. You are the pilot in command and nothing else is lawful. Before a flight we read the rules you already wrote — your manual, and the conditions attached to your approval — and say what we find: nothing here says no, or this is against a rule you wrote, quoted in your own words with its section and page, or I can’t tell, naming the fact we could not establish.

Whatever you decide, it is recorded — the rule, the situation with each fact’s provenance, who decided, and the hash of the version of the manual that was in force that day. Every other system logs that a flight happened. This logs which written rule authorised it, in the words that were in force.

What a record does and does not establish

A record establishes what was decided, by which rule, under which version of policy, by whom, and that it has not changed since it was sealed. It does not establish that the underlying facts were accurate. The verifier is open source, runs offline, and needs nothing from us — a record only readable inside our software would not be a record.

And underneath it, the part that remembers being fooled.

Reflex is a second detection channel on the same camera input. It shares no model, no training pipeline and no inference runtime with your perception stack, reaches its own answer by a different route, and flags the frames where the two disagree. That is the catch, on that aircraft, in that frame.

Then the second half. Every aircraft in your fleet learns alone. Something gets past the perception stack on one airframe, and the next airframe falls for it too, and the one after that. Fixing it means a labelling cycle and a retrain, which is not something you can run in the field.

We are not the brain of the aircraft. We are the part that remembers being fooled.

Before any of that: your manual and the conditions attached to each approval were written across years by different people, and nobody has read them end to end. Paste twenty of your rules and we will tell you which two cannot both hold. No account, about a minute.

Check your rules — free
The compute
An NVIDIA Jetson Orin Nano developer kit on the bench, the board the reflex layer runs on

NVIDIA Jetson Orin Nano. Reflex is built for the Jetson Orin that flies on your aircraft.

Where it sits
COMPUTE BOARD your perception stack reflex COMPUTE BOARD your perception stack reflex one small message no central node Reflex raises what the stack missed, on that aircraft. Then the rest of the fleet has it too.

What it is

Catches what your stack missed.
Shares it without a retraining cycle.

01 — A SECOND CHANNEL

Not a better detector. A different one.

Its own model, its own training pipeline, its own inference runtime. Nothing it does depends on your perception network being right. It reads the same input, reaches its own answer, and raises the frames where the two disagree.

02 — WHAT COMES OUT

Graded, not a per-frame alarm

One disagreeing frame raises nothing. The response is graded and dwells before it escalates, so what reaches your autonomy is a considered signal rather than a stream of noise. What the aircraft then does about it stays with your flight stack.

03 — IT RUNS ON JETSON

On the board you already buy

NVIDIA Jetson Orin, including the National Defense Authorization Act (NDAA) compliant carrier boards you already source. It installs onto the image you already flash, as its own process. Your detector and your flight controller are untouched, and swapping detectors later does not involve us.

NVIDIA Inception Program

Univault Technologies LLC is a member of the NVIDIA Inception program.

04 — ONE TEACHES THE REST

One message, and it is small

When one aircraft meets something new, what it learned travels to the others as a single small message. No call back to a central node, and nothing in your flight-critical software changes when it arrives. If the link is down, it waits.

Where the security layer fits. Every message is signed by the aircraft that produced it, and an aircraft ignores anything it cannot verify. Otherwise one compromised airframe teaches the rest something false. Signing is a separate module, so a program that is required to use a specific cryptographic implementation points it at that one instead of ours.

Who it is for

Builders who already
chose Jetson.

If you build small uncrewed aircraft on NVIDIA Jetson, ship to defence or public safety, and your fleet keeps meeting the same problem one airframe at a time, this is aimed at you. If you fly a different compute stack, tell us anyway — it changes what we build next.

Register interest

Tell us what is
getting past you.

We are pricing the first release now, and we would like to quote you. Tell us what your fleet keeps running into and we will come back with a quote for the first release, the board it ships on, and a straight answer on whether it helps you or not. The founder reads every one of these and replies to it.

Goes straight to the founder.