Our Tech | SpectralSense

02 / WHY NOW

What changed in the last year

For twenty years, serious computer vision meant a camera that sent pictures somewhere else: a server in a closet or a data center in the cloud. The camera saw. Something far away decided. That is why the systems in the field today keep video, cost a fortune to run, and stop working when the network does.

Three things arrived at once. Vision processors that run a full detection model on a few watts. Accelerators with their own memory, the size of a credit card, that run a small language model. And stereo depth modules at commodity prices. Put them together and a single unit at the door can see in three dimensions, run the model, explain what it saw, and keep the record, with no server and no cloud in the loop.

We are the ones who put them together. Everything we ship runs on that stack.

03 / HOW IT WORKS

From the lens to the decision

Every SpectralSense system follows the same five steps. The sensors and the rules change by product. The pipeline does not.

SENSEStereo cameras give us depth as well as color. Spectral sensors read the actual color of a material. A passive cue can wake the camera.Depth is the information legacy systems throw away. A flat photo is never counted as a person.

BASELINEComputer vision learns what normal looks like at that camera: the gear, the light, the traffic, by hour and by role.A few reference photos, taken on site. No training run.

RULESThe client sets the rules. A written policy becomes a profile the device can check against.Clothing, objects, who belongs, when, which direction.

INFERModels run on the device, on a dedicated vision processor. A verdict takes a few milliseconds to a second.Quantized models compiled to the hardware. The host CPU is not in the loop.

RECORDEach decision is written to a hash-chained log on the unit and synced when there is a network.About 60 bytes per event. Time, place, rule, verdict. No image.

Inference runs in two speeds. A fast detector handles every frame. When it is not sure, a second model on the accelerator looks at the same frame and explains in plain words what it sees and why it does not match the rule. A person makes the final call, and only that confirmation changes what the device looks for. The second model never trains the first on its own, so the system gets sharper at its own camera without making things up.

The same code runs on the edge unit, on a phone, and in our cloud, and all three give identical answers on the same input. That is what lets a customer start on a phone and move to a fixed unit without rebuilding anything.

04 / THE STACK

Where we sit

We do not make cameras and we do not make chips. We make the layer that turns commodity sensors and commodity silicon into a machine that can be trusted to decide on its own. Here is the whole stack, bottom to top.

SENSORSGlobal-shutter stereo pair with active infrared, 12MP color, multi-channel spectral sensor. All off the shelf.

SILICONArm system-on-module, a dedicated vision processor at the sensor, and an AI accelerator with its own memory for the second model. The whole unit draws less power than a light bulb.

PERCEPTIONOurs. Depth-masked detection, reference-set matching against the client's own gear, worn-versus-carried logic, spoof rejection, and a photo-to-numbers reduction that keeps 160 numbers and discards the image.

RULES ENGINEOurs. The client's written policy, turned into a signed profile the device checks against. Versioned, so the record shows which rule was in force and who published it.

RECORDOurs. An append-only, hash-chained log written on the unit before anything leaves it. Open formats out: events, CSV, PDF. Signed over-the-air updates in.

CLOUDRule authoring, fleet updates, dashboards and exports, sign-in through the customer's own identity provider. The cloud is for authoring and reporting. The decision never goes there.

Because the three middle layers are ours and the two ends are commodity, we can move to new sensors and new silicon as they get cheaper without rewriting the product. That is the whole bet. The hardware under us will keep getting better every year, and we ride it.

05 / LEGACY SYSTEMS

Why the current systems cannot follow

The systems in the field today were built before any of this was possible, and their business depends on the thing we took out: pictures of people, stored somewhere else.

	Where it decides	What it keeps	Sees depth	Works offline	Liability it creates
CLOUD VIDEO ANALYTICS	a server, after the fact	clips, thumbnails, faces	no	no	biometric, breach, discovery
ON-PREM GPU BOX	a server in a closet	video, on your hardware	rarely	yes, if the closet is up	same, now it is your problem
WEBCAM AND A LAPTOP	a laptop, in 2D	whatever the app saves	no	sometimes	fooled by a photo
SPECTRALSENSE	the device, in under a second	about 60 bytes, hashed	yes	yes, by default	none. no image exists

A video vendor cannot get here by improving. They would have to stop storing video, and video review is their business.

06 / WHERE IT RUNS

One pipeline, many products

The pipeline does not care what it is looking at. Change the sensors and the rules and it is a different product. These are the ones we have built so far.

WorkCheckChecks what a person is wearing against the rule for that door, decides in about a second, and keeps a record that holds up. The product

LookSenseThe same reference matching on a phone, for people deciding what to wear. The consumer side of the pipeline. More

The PrismA handheld spectral sensor that reads the true color of a fabric or a material, independent of the light in the room. More

TripwireA passive cue wakes a camera, and the camera puts a 3D track on a small, fast target. The same edge inference, pointed at the yard instead of a door. Patent pending.

Defense and governmentRecords of compliance at the flight line, the depot, and the base gate, without touching anyone’s identity. More

NextFood plants, cleanrooms, pharma gowning, franchise uniform checks, venues, training stations. Anywhere a camera has to decide on the spot and prove it later.

07 / WHY US

We have done this before

This is not our first vision system. Our founder has spent thirty years on optics, computer vision, and camera calibration, and the team behind him built some of the largest real-time visualization systems ever installed. We calibrated multi-projector domes. We built multi-camera touch systems that predicted where a finger would land before it got there. We wrote real-time rendering engines and camera-to-projector registration that had to be right to the pixel in front of a live audience.

Edge AI is the same discipline with the computer moved next to the lens. The hard parts are the ones we have always done: knowing what a camera can and cannot see, making the math run in the time you have, and building a system a customer can trust when nobody is watching it.

That is why this works now, and why it is ours to build. The hardware finally caught up to what we already knew how to do.