Capabilities

The Work
We Do

We design the sensor, test it, and write the software that runs on it.

Earth: NASA · beam and footprint are an illustration · credits

01

EO/IR system design and integration

Custom Imagers, Visible to LWIR

We design the imager and its processing chain together, from the band and the focal plane to the packaging that gets it onto the platform.

  • Band selection, optical and focal plane trades
  • Real-time image processing, including plume imagery
  • Precision target tracking
  • Field calibration procedures written for the operator
A low-light visible frame at night: a white pickup on a driveway in front of a large tree, with house lights in the distance. The same scene in long-wave infrared, white-hot: the tree canopy and the truck's engine and tires bright against a dark sky and cooler ground. The visible and infrared frames blended into one image. Visible · low light LWIR · 7–14 µm
Drag the divider

Two bands, one scene. A low-light visible frame and a Velox LWIR frame, registered to one geometry. The registration is global, so near and far objects keep a little parallax. Images: Velox Research.

Multispectral coverage 0.4–14 µm · select a band

Visible · 0.4 to 0.7 µm

Reflected daylight, the same thing the eye sees. Highest spatial detail of any band, but it needs light and clear air.

Best for: daylight imaging · high-resolution reconnaissance · visual identification

02

Synthetic IR imagery

Synthetic IR, Scored Against Real Frames

Our pipeline generates synthetic infrared imagery on ordinary CPUs, and scored against our own real collections it has closed the gap between synthetic and real IR.

A long-wave infrared frame from a Velox sensor: trees and ground in white-hot thermal grayscale, with the sensor's own green detection box around a warm target at the base of the tree line.

Real. A daytime LWIR frame from a Velox sensor, the kind of collection the synthetic output is scored against. The detection box is the sensor's own.

03

Sensor test and evaluation

Bench, Range and Data Reduction

We characterize a sensor on the bench, take it to the range with instrumented ground truth, and reduce the data. Our own hardware and other people's.

  • Radiometric and geometric calibration
  • Environmental and field test
  • Signature collection in IR, RF and acoustic
An imaging spectrometer instrument mounted inside a cylindrical thermal-vacuum test chamber, wrapped in reflective thermal blanketing.

An imaging spectrometer in a thermal-vacuum chamber. NASA/JPL-Caltech.

04

Aerospace systems engineering

Hypersonics, Technical Collections, Missile Defense

Systems engineering from concept through algorithms to hardware-in-the-loop, for technical collections and missile defense.

  • Collection architectures
  • Algorithm development and DSP against representative data
  • Analysis of high-speed cruise missiles and projectiles
Schlieren image of two supersonic T-38 aircraft in formation, with the shock waves from each aircraft visible as sharp bands merging behind them.

Air-to-air schlieren of two supersonic T-38s. NASA, 2018.

20 KM40 KM60 KM IR TRACKER · SITE 1 SITE 2 BASELINE 91 KM · SURVEYED PLANCK · LEADING EDGE · 0.5–14 µm SWIRMWIRLWIR HYPERSONIC TRACK · IR SIGNATURE T₀ = T∞ (1 + ½(γ−1) M²) · λₚₑₐₖ = 2898 µm·K / T · γ = 1.4 ILLUSTRATIVE · 1 PX = 0.5 KM TRACKER TRACKING T+00.0 s MACH—ALTITUDE—FIX—RANGE S1 / S2—ELEV S1 / S2—STAGNATION T₀—LEADING EDGE T—λ PEAK · WIEN—BAND— IRRADIANCE AT SENSOR ∝ T⁴ / R² S1 S2 DETECTION THRESHOLD T₀: IDEAL GAS, γ = 1.4 · EDGE T: SIMPLEMODEL · NOT FLIGHT DATA
A glide vehicle at Mach 5 to 9 heats its leading edge past a thousand kelvin, and where that emission peaks decides whether SWIR or MWIR tracks it. Two sites on a surveyed baseline turn bearings into range. Demo geometry and a simplified temperature model.

05

Low-power edge & embedded sensors

Embedded Sensors That Watch for Months

Hardware sized to what a field unit can carry. It processes data where it is collected and wakes the radio only to send results.

  • Analog, digital and power design
  • RTOS and Linux targets, including PetaLinux and Yocto
  • Ruggedization and shock hardening
Top-down photo of a Xilinx Zynq-7000 development board: the SoC in the centre, memory beside it, and the connectors along the edge.

A Zynq-7000 node on the bench. Board photo: Wikimedia Commons, CC0.

06

Autonomous systems for ISR

Tracking and Autonomy at the Sensor

Tracks fused from more than one sensor, and autonomy that runs on the sensor itself so the system can act without a person in every loop.

  • Stereoscopic 3D tracking
  • Multi-source fusion and track association
  • Autonomy written for the specific sensor
STEREOSCOPIC TRACK GEOMETRY DEMO · 1 UNIT = 4 M BASELINE · 1.68 KM · SURVEYED S1 S2 S1 BEARING062.4° S2 BEARING118.9° RANGE FROM S11.83 km ALTITUDE412 m CLOCKT+04.2 s TWO BEARINGS FROM A KNOWN BASELINE GIVE A 3D FIX WITH NO RANGEFINDER ON EITHER SENSOR
Two lines of sight from a surveyed baseline cross at one point. Demo geometry, not recorded data.

07

Space vehicle digital twins

The Flight Computer on the Bench

The real flight binary runs unmodified on an emulated processor, closed-loop with a modeled vehicle, so faults, bad commands and attacks get tried on the ground first.

  • Emulated flight computers that run the flight binary as is
  • Regression test, operator training and cyber evaluation on one twin
  • A laptop for most work, Xilinx-class FPGA boards when timing matters
A Xilinx Zynq-7030 system-on-module: the Zynq in the centre, memory chips around it and a bar-coded label at the bottom.

A Zynq-7030 module, the class of part the twin runs on when timing matters. Wikimedia Commons, CC0.

08

On-device AI / ML

Small Language Models That Run Offline and on Mobile Devices

Speech, text, images and video processed on the phone, laptop or edge box that collected them, by the newest open models (Gemma, Meta's NLLB, Whisper) quantized to fit the device. No connection back is assumed, and nothing leaves the device.

One Arabic message becomes text by transcription (from audio) or OCR (from a photo), then fans out to a transliteration in Latin letters, an English translation, a short summary, and extracted fields for when, where and what to bring. Everything stays on the device.
Transcription and OCR change the form, transliteration the script, translation the language, and summary and extraction the length. Whisper for speech and NLLB for translation on the phone, with language packs loaded before the mission.
Fig. 01   Detect, track, type  ·  loop

Video analytics on Frigate, an open-source NVR: a detector and tracker box each object, then a second model names the type. Footage: U.S. Marine Corps and U.S. Army via DVIDS, public domain. The appearance of U.S. Department of War (DoW) visual information does not imply or constitute DoW endorsement.

Three people walking down a city street. Each face has a detection box, five keypoints and a landmark overlay. Person 1 is labeled MATCH 0.84 against gallery template 001; persons 2 and 3 are labeled NO MATCH, NEW ID.

Faces found and matched on the device. Real detector, landmark and match output on a street clip. Person 1 was enrolled three seconds earlier, and a person confirms any match that matters. Footage: Pexels.

  • Edge AI and ATR on embedded MCUs, making decisions at the edge under latency and bandwidth limits
  • Offline transcription, translation and OCR as Android, iOS and desktop apps
  • Detection, tracking and face recognition on a single edge box
  • A local Gemma-class model and search over everything collected that day, with sources shown
  • Delivered as installable software, updated by signed bundles, measured on the target device

09

ICS / OT security

Security for Controllers That Cannot Stop

A smaller line of work: security architecture and assessment for the PLCs and edge devices that run a physical process and cannot be taken offline to patch.

  • Segmentation and monitoring that respect real-time limits
  • Modern security on legacy systems that were never designed for it
A Siemens SIMATIC S7-1500 PLC rack: power module, CPU with its small display, and a row of digital and analog I/O modules.

A Siemens S7-1500 rack. Photo: Palatinatian, Wikimedia Commons, CC BY 3.0.

Need One of These on a Program?

Tell us the problem and the constraints. We will tell you whether we are the right shop for it.

Contact Velox