Airman VisionAirman
Vision
See threats. Keep secrets.
Private, on premises AI for security, vision and cyber defense. One software platform, running on NVIDIA GPUs inside the customer's own walls.
Private AI and Vision prototypes running today on local NVIDIA hardware
The problem
Security teams want AI.
They cannot ship their data to someone else's cloud.
Cameras cry wolf
Most video analytics alerts are false alarms. Guards learn to ignore them, and the real threat gets lost in the pile.
Cloud AI leaks
Law offices, schools, municipalities and defense suppliers cannot send sensitive footage, documents or network logs to a public model.
Compliance is a burden
CMMC and rules like it demand evidence and monitoring that small teams simply cannot staff.
The platform
One platform. Three modules.
The same software stack runs on every size of system: CUDA, TensorRT, Triton, DeepStream, Morpheus, Nemotron models, private retrieval and disk encryption. Upgrade the GPU, not the software.
Private AI
A local LLM workspace with private retrieval over the customer's own documents. No data leaves the building.
How it works →Vision
Firearm, aggression, fall, loitering, and fire and smoke detection, with a vision language model second check on every alert.
How it works →Cyber defense
A network defense agent and a CMMC compliance agent on NVIDIA Morpheus, reading traffic locally.
How it works →The second check
Every alert is re read by a vision language model before a human sees it.
Cheap first, smart second
The fast detector runs on every frame. The slower VLM only runs on flagged clips, so cost stays low.
Auditable
Each decision keeps the clip, the prompt, the VLM answer and the operator action for later review.
Architecture
Detect at the edge.
Reason in the core.
Jetson units at the camera run the detector on the live stream. An Airman Vision server on site handles the second check, footage search, analytics and agents. A DGX or the customer's own private cloud is optional, for heavy VLM work, fleet analytics and retraining.
Raw video stays on the edge or site box. Only alerts, metadata and clips the customer approves move upstream.
Footage search
Ask your cameras a question.
Example queries
“Person in a red jacket near the loading dock after 6 pm.”
“Show every vehicle that stopped at gate 2 last night.”
Planned. Built on the NVIDIA video search and summarization blueprint. Clips are embedded, captioned and indexed on the site box, and nothing is sent to an outside service.
Follow the work
Prototype. Benchmark. Pilot. Ship.
Every figure we publish will come from our own runs on named hardware, with scripts alongside. Researchers, NVIDIA developers and security practitioners are welcome to compare notes.
[email protected]