Velox Labs

Applied R&D

We build the kit, take it outside, and bring the data back to the model. Work here starts before anyone has written a requirement for it. When a number on this page is a target and not a measurement, it says so.

01 · Bench

Kit we build, integrate and test.

Hardware We Build On

Airframes, edge AI kits, cameras and prototypes, next to Jetson modules, Zynq UltraScale+ boards and workstation GPUs. If it is going to fly or sit in a field for a month, it gets built and broken here first.

A twin-boom VTOL fixed-wing UAV laid out on a conference table during integration, lift motors on the booms and propellers set beside them.

A twin-boom VTOL airframe during integration. Photo: Velox Research.

An STMicroelectronics STM32N6 Discovery kit on a Velox mat: a five-inch touch display board with the IMX335 camera module on its own small board above it.
Two cameras on a Velox mat: a grey Pixelink machine-vision camera with a Navitar lens, and a red camera body with a long lens, a long-pass filter and a T-mount adapter.
A prototype on a laser-cut glowing green Velox panel under UV light: an 18650 battery pack, a metal pushbutton, an electromagnet, a time-of-flight distance sensor, a MOSFET driver and a Wi-Fi microcontroller, wired together.

An STM32N6 edge AI kit with its camera, two camera bodies with one behind a long-pass filter, and a battery prototype that fires an electromagnet from a range reading. Photos: Velox Research.

02 · Field

Local AI for places with no link back.

Tactical Edge Compute

Out there, nothing has a data center behind it. We build air-gapped AI stacks on commercial hardware and run the exploitation where the data is collected.

The same kit goes to the range. The array on the right is tracking a quadcopter by sound, and the case below is collecting the magnetic signatures of ground vehicles.

Acoustic tracking: a four-microphone array on a red Velox-printed cross mount and tripod in sandy grass, with a small quadcopter far off above the pine tree line.

A four-microphone array on a Velox-printed mount, with a quadcopter at standoff range. Image: Velox Research.

A

Multi-Modal Ingestion, Fully Offline

OCR, summarization and cross-referencing with the network interface down. No cloud inference, no telemetry.

Built and tested on: NVIDIA Jetson AGX Orin  /  tactical desktop rigs

B

Semantic Edge NVR

An open-source NVR with a local vector database, so an operator can search footage by describing what they want to see.

Stack: FOSS NVR  /  Qdrant vector database  /  local embedding + summarization models

Magnetic collection of ground vehicle signatures: a black instrument case with sensor units on top in tall grass in the foreground, and an olive-drab box truck passing on the far side of the field, with pine trees and a summer sky behind.

Magnetic collection of ground vehicle signatures. Image: Velox Research.

03 · Models

Model the field before building hardware.

What Comes Back Feeds the Model

Field data feeds the models, and the models decide what gets built next. Electromagnetic models map the field around a source, so sensor placement and standoff are worked out before any hardware exists, and bad geometries get thrown out in software where it is cheap.

Fig. 01   Field model  ·  loop

Streamlines around a spherical source, colored by |B| on a 10−7 T scale.

The same goes for aerothermal work. Our CFD runs on real vehicles. This one is a reindeer at hypersonic speed, for fun, and the gas still hits about 1,500 K at the nose.

Computational fluid dynamics cut plot of fluid temperature around a reindeer model in high-speed flow. The gas is hottest, near 1,500 kelvin, at the nose and chest, with a hot wake trailing behind; a colour bar runs from 50 to 1,500 kelvin.

Fluid temperature, 50 to 1,500 K. Colour scale re-rendered for the page.

04 · Next

Where the next concept gets built.

Quantum Magnetometers

The current bet. We characterize candidate quantum magnetometers against the modeled field, so we know what a sensor can resolve before it leaves the bench, and we are building our own concept ahead of any requirement. What survives moves into program work. What does not gets dropped while dropping it is still cheap.

A magnetometer sensor head hanging from a spacecraft boom in a cleanroom, with gloved hands working on it.

Magnetometer integration on NASA's Psyche spacecraft, shown as an example. NASA/JPL-Caltech.

Have a Hard Problem?

Labs work starts as a conversation with a program office or a prime that has one.

Contact Velox