PARAGON REFLEX ™

A reflex layer and a language layer for autonomous machines.

Catch what your detector missed. Then the whole fleet has it.

One process beside the perception stack you already run, sharing no model, no training pipeline and no runtime with it.

One network decides whether there is a hazard, and nothing downstream can tell when it is wrong. We read the same camera and lidar input, reach our own answer, and raise the frames where the two disagree. When one unit meets something new, the rest inherit it as one small signed message: nothing is retrained, nothing is uploaded, no server sits in the middle. Where a person, a floor or a fleet needs a call it can sign, the Reflex Planning LLM reads the feeds, picks from your written rules, and says when the evidence isn't there.

Own model| Own runtime| Says when it is unsure| Sized for the safety controller you already carry
Reserve a kit Request a key Read the paper

Tier-1 suppliers and OEMs: start an evaluation on your own logs

Machines that learn with you. AI that keeps you the author.

A product line of Univault Technologies. Patent-pending.

Camera and lidar input | Own model | Own training pipeline | Own runtime | NVIDIA Jetson Orin | Automotive Safety Integrity Level D decomposition target | Patent-pending

Two units, and you

Two boards on a desk.
The first taste of the whole idea.

Two Jetson boards, a camera on each, the reflex layer already beside your detector. One board meets something new; the other already knows. You are the third learner, and the record of what all three of you learned is yours. That is the Reflex Dev Kit, our first product on purpose.

Dev kits do not scale; builders do. The kit is the on-ramp to a per-unit licence on the compute you already ship.

BOARD A your detector reflex BOARD B your detector reflex one small signed message THE NODE what was it? raccoon Board A meets something new and raises it. Board B has it before it looks. The node asks; you answer.
FULL KIT
$999
one kit, everything in the box
  • 2 × NVIDIA Jetson Orin Nano developer kit
  • 2 × camera, brought up on the layer
  • 2 × storage, pre-flashed, ready to boot
  • The reflex layer, installed beside your detector
  • Two-board licence, signed transfer
  • The record, yours, as plain files
Reserve the full kit
BRING YOUR OWN BOARDS
$449
you already own two Jetson Orin boards
  • 2 × camera, brought up on the layer
  • 2 × storage, pre-flashed for your boards
  • The reflex layer, installed beside your detector
  • Two-board licence, signed transfer
  • The record, yours, as plain files
  • 90 days of Reflex Planning, 10,000 answers
  • Ships within 10 business days · full refund before shipment
Reserve the BYO kit
DEFENCE TIER
By quote
NDAA-compliant Orin Nano, US board vendor
  • You buy the compliant boards from the vendor
  • We put the layer on them, remotely
  • Same kit, same record, same licence
  • Selectable cryptographic backend
Ask for a quote
REFLEX PLANNING LLM
$0.01 an answer
first 1,000 free
  • The language layer for the machine, not a chatbot: what you pay for is what the labs do not sell
  • Talks with the person, picks one of your numbered actions, reports to the record
  • Every answer carries a line: confident, or not; not confident = the unit does not move
  • 1,000 answers free; kit tenants get 90 days
  • Fleets: we host it for you, or it runs in your account; by quote
Request a key

Reserving holds a place and charges nothing; we confirm price, board and ship window by email. Under a thousand dollars on purpose: a lab card, not a purchase order. Everything in the kit →

One teaches the rest

Fooled once,
not once per unit.

When one unit meets something its detector got wrong, it learns the catch right there and passes it to the others as one small signed message, so they already know it before they meet it. Nothing is retrained, nothing is uploaded, and no server sits in the middle.

Every unit writes down what it met, what it learned and what it passed on, in a form an outside assessor can read. The record is yours. It is also the start of a certification path: approve the procedure once, instead of re-approving every change.

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 unit. Then the rest of the fleet has it too.

Two products, one machine

The kit is the body.
The Planning LLM is the layer between complex signals and precise human actions.

A reflex arc does not consult anything before it acts; that is the kit, and that is why it is the body. But a unit in a barn, a drone cell under a stadium, an operations floor reading hundreds of alerts a shift — all of them have to turn a flood of feeds into one call, by written rules, in seconds, and defend it afterwards. Reading fast is the half current AI does. The other half — follow the rules, hold when the evidence isn't there, write it down — is why today's chatbot LLMs are not suitable for many embodied AI tasks. We built Reflex Planning LLM for that half, and we measure it on the number a fleet or a floor actually pays: what a correct answer costs.

Why not the LLM you already have

WORDSIt gives you words. A body needs an action it can trust, from a list you wrote.
UNSUREIt does not tell you when it is guessing, so a wrong answer becomes a wrong movement.
COSTNobody measures it on what a correct answer costs per unit, per turn, all day; a fleet pays that number every day.
LEARNINGWhat it picks up from a person is not named, not bounded and not revocable. Ours is all three, or it is not learned.
TIMEFlagship models answer in seven seconds to over a minute. A person standing in front of a unit gives you about two.
TRACEEvery call can be traced back: the feeds it saw, the rule it applied, the line it picked or held, who confirmed it, when — and revoked with one signed line. A chatbot leaves a transcript; nobody can say afterwards which rule produced which action, because there was no rule, only a prompt.
REFLEX PLANNING LLM — THE OTHER HALF
READEvery feed, your written rules, the standing orders, the question in plain words.
DECIDEOne of your numbered lines, or hold; what the evidence establishes and what it does not.
RECORDWhat was seen, what was called, who confirmed it, when; revocable; readable by an assessor.
See the Planning LLM →

We publish what a correct answer costs, measured against the top labs on how well the model follows human rules, and give partners the template to measure it against their own operations manual. The paper →

Where it runs

Built and measured on road autonomy. Now on NVIDIA Jetson, the board your aircraft, robots and cameras already run.

An NVIDIA Jetson Orin Nano developer kit, the board the reflex layer runs on

The same second channel, on the Jetson that small uncrewed aircraft already carry. It reads the same input as the perception stack, reaches its own answer, and raises the frames where the two disagree. And because a fleet is many aircraft rather than one, what a single airframe catches then travels to the rest instead of being met again by each of them.

Reflex for uncrewed aircraft

Why this matters

The fleet finds it Tuesday.
The fleet gets the fix
two quarters later.

A vehicle meets something your perception stack does not flag. Turning that one event into fleet-wide behavior means a labelling cycle, a retrain, a full revalidation of a network that changed underneath you, and a software-update file that has to be re-opened.

TODAY Tuesday label retrain revalidate reopen the update file two quarters later WITH THE LAYER Tuesday the rest of the fleet has it same day. Weights never move. The bottleneck was never perception. It was the correction loop.
Why it is always two more years

Autonomy is usually described as a perception problem. The bottleneck is the correction loop.

Label it, change the model, prove nothing else broke. Once the weights move, your existing evidence describes a system you no longer ship, so the safety case gets re-argued and, in type-approval markets, the update file gets re-opened.

Then the fleet finds four more.

Almost none of this is bounded by compute, and compute is the part you can buy more of. The bound is the people who write the safety argument and the independent assessors who sign it, and that capacity does not grow because you need it to.

The date moves because the loop that corrects a fault runs slower than the loop that finds one. You can tell where the industry thinks the constraint is by what it is paying for. Waymo is spending to compress the time it takes to open a new city. Mobileye built a map that updates itself off the fleet. Wayve is betting on not needing a map. None of that is a bet on better driving. All of it is a bet on making the next change cost less than the last one.

An independent monitor on your perception channel, built so that keeping it current costs less than rebuilding it.

When the backup is another network
  • ✕
    Every fix is a retrain
    New weights mean your argued safety case describes a system you no longer ship, and your assessor has to form a view on the change.
  • ✕
    Correlated failure modes
    Two networks trained on the same conventions tend to be wrong about the same scenes, and share the accelerator besides.
  • ✕
    The scarce resource is people
    One retrain spends the same review capacity whether it fixes one scenario or forty, and that capacity is booked months out.
  • ✕
    Couples you to a vendor's roadmap
    The backup constrains which detector you are allowed to choose next.
With a reflex layer
  • ✓
    Updates stay small
    Keeping the channel current does not move your perception network’s weights, so the safety argument you already made for that network still describes what you ship.
  • ✓
    Sized for the other processor
    Kilobytes of state and a fixed execution cost, so independence can be physical.
  • ✓
    Not a network
    Its own model, its own training pipeline, its own inference runtime. Nothing it does depends on yours being correct.
  • ✓
    Does not constrain your detector
    It does not touch your model or your choice of detector, so you keep choosing.

The evaluation

You set the bar.
Then we run it on your data.

The only quantity worth anything to you is how much of your own stack’s residual this channel recovers on your own fleet’s data. That number does not exist in a public benchmark and we are not going to manufacture a proxy for it. So we do not lead with numbers. We run the test you specify, on a bench, with your logs.

YOUR LOGS, REPLAYED your detector reflex THE BAR YOU SET ! ! ! Where the two agree, nothing happens. Where they disagree, the frame is raised. Your team scores it against the criteria your team wrote.
You write the criteria

Pass and fail are defined by your team, in writing, before we see any of your data. Including the scoring convention and the false-alarm budget you actually have to live inside.

Your logs, your labels

A replay on your recorded data against your own detector. No integration, no vehicle time, and nothing installed on anything that moves.

You keep the verdict

Your team scores it against the criteria your team wrote, and keeps the result either way. We see the same report you do.

How we release results. A result about a safety layer is only readable with its conditions attached: which detector it was measured against, the false-alarm rate both sides were held to, the distance band, and the definition of a miss. Those four travel with every number we release, under a mutual non-disclosure agreement.

Where it runs, and what it is for

Three things to check before you talk to us.

None of these needs a benchmark to evaluate. Two of them you can check against hardware you already own.

01 — THE BOUNDARY Independent of your perception software. Check it on your own logs. 02 — THE BUDGET Sized for the safety controller you already carry. Check it on hardware you own. 03 — THE UPDATE PATH A smaller thing to argue than shipping a different network. Ask your assessor.
01 — THE BOUNDARY

Model failure, not sensor failure

The reflex layer has its own model, its own training pipeline, and its own inference runtime. Nothing it does depends on the perception network being correct. It is not sensor-independent: both channels read the same sweep, so a blinded sensor takes both down together. The independence is from the perception software, and that is the failure family it is there for.

02 — THE BUDGET

Sized for the safety island

Two kilobytes of state and a fixed-cost comparison against it. No floating-point network inference, no dynamic allocation, and control flow that does not branch on the data. That is a budget we are targeting at the lockstep microcontroller already in the vehicle rather than the accelerator inside the main system-on-chip. The port to that microcontroller is in build.

03 — THE UPDATE PATH

The update path

When the fleet finds a hazard pattern the car did not have, updating the channel for it is designed to be a smaller thing to argue under an update-approval regime than shipping a different network. No assessor has given an opinion on that yet.

What this is, and is not, today. At its current false-alarm rate the reflex layer is a monitoring, arming, and event-capture channel. It is not a brake authority. The false-alarm rate is the gate, and closing it is the program we are running.

Six markets, one layer

Six markets already pay
to catch the miss.

We do not build the security company, the farm robot, the drone service, the safety system or the backyard camera. Builders do, and they already exist in every market below. What none of them has is a layer that catches what their detector missed, passes it to the rest of the fleet without a retrain, and writes down how it learned.

Site and venue security Farm and ranch Drone services Industrial safety Field robots Home and pets ONE LAYER Builders run the business. The layer, the learning and the record ride inside every unit.
Site and venue security
01
Site and venue security

Physical security is a roughly $120 billion a year market, and video surveillance is more than half of it. Every camera on a site already produces frames a detector sometimes reads wrong, and every miss gets reviewed afterwards.

Source: MarketsandMarkets, Physical Security Market 2025-2030; Video Surveillance Market 2025-2031.

What the layer does here. A second channel on the cameras already installed catches what the first one missed, and the next site inherits it. The record is what the after-action review reads.

How the kit puts you in it. Integrators start with one kit on two cameras, then license the layer per camera.

Our role: the layer, the learning, the record — inside every unit. Your role: the business.

Farm and ranch
02
Farm and ranch

The FAO estimates up to 40 percent of global crop production is lost to pests each year, and plant diseases cost the world economy more than $220 billion annually.

Source: FAO, Scientific review of the impact of climate change on plant pests, 2021.

What the layer does here. Field units with no connectivity, every farm different, one unit teaching the rest: predators at the fence, the gate left open, the row that looks wrong this week.

How the kit puts you in it. Ag robotics builders start with one kit; the farmer is their customer, the layer rides on every unit they ship.

Our role: the layer, the learning, the record — inside every unit. Your role: the business.

Drone services
03
Drone services

Drone services, inspection, public safety and delivery, is a market analysts put between $16 and $32 billion in 2025, depending on who counts, and growing at double digits.

Source: Mordor Intelligence, Fortune Business Insights, Precedence Research, 2025 estimates.

What the layer does here. What one aircraft gets fooled by, the rest of the fleet should only be fooled by once. The layer is built for the Jetson Orin that already flies.

How the kit puts you in it. Drone builders on Jetson start with one kit on the bench, then a fleet licence on the board they already buy.

Our role: the layer, the learning, the record — inside every unit. Your role: the business.

Industrial and warehouse safety
04
Industrial and warehouse safety

Work injuries cost the United States an estimated $176.5 billion in 2023, before a single fine or lost contract.

Source: National Safety Council, Work Injury Costs, 2023.

What the layer does here. Zones, forklifts, people where they should not be: the incumbent system catches most of it. The layer catches what it missed, and the next site starts with that already learned.

How the kit puts you in it. Safety integrators start with one kit at one dock door, then license per site.

Our role: the layer, the learning, the record — inside every unit. Your role: the business.

Field robots
05
Field robots

The machines are already in the field. Harvesters, sprayers and scouts carry cameras and compute today, and each one learns alone.

What the layer does here. One unit meets the thing the detector never saw; the rest of the fleet has it before the next pass. No retrain, no upload, no waiting for the season to end.

How the kit puts you in it. Robot builders start with one kit on the bench; the layer ships as a software option on the compute they already use.

Our role: the layer, the learning, the record — inside every unit. Your role: the business.

Home, pets and the backyard
06
Home, pets and the backyard

Americans spent $158 billion on their pets in 2025, and 95 million households own at least one. The camera on the porch already sees the raccoon; it just calls it a cat.

Source: American Pet Products Association, State of the Industry, 2026.

What the layer does here. Two units and a node on the kitchen table. The owner answers one question and the yard learns. This is the consumer volume, and the on-ramp for everyone above.

How the kit puts you in it. Product builders start with one kit; the layer ships inside the feeder, the coop camera, the yard unit they sell.

Our role: the layer, the learning, the record — inside every unit. Your role: the business.

Start on a desk. Scale when the record says so.

Builders start with two boards and a key. Tier-1 suppliers, OEMs, robotics platforms and eVTOL (electric vertical take-off and landing) programs start with an evaluation on their own logs, against a pass bar they set. The paper is the measuring stick for both.

Reserve a kit Request a key Read the paper

Tier-1 suppliers and OEMs: start an evaluation.

Press resources