Research

Benchmark, then prove it.

Security AI makes a lot of claims. Airman Vision plans to back every one of them with numbers anyone can check.

Planned research with Dr. Wang

Four model families. One fair test.

Compare the main approaches to computer vision head to head on NVIDIA GPUs, with and without TensorRT optimization.

YOLO

Fast single pass object detectors, the workhorse of real time video.

CNNs

Convolutional networks, the proven backbone of image recognition.

Vision transformers

Attention based models that see the whole frame at once.

Multimodal AI

Models that combine vision and language, like the VLM reviewer.

Test bench

NVIDIA GPU benchmarks

Each family runs on the same footage and the same hardware, once as trained and once after TensorRT optimization, so the speed gain from optimization is measured rather than assumed.

Publish the evidence

Accuracy / False alarm rate / Speed

Accuracy alone is not enough. A detector that catches every gun but also flags every umbrella is useless to a security team. That is why false alarm rate gets equal billing, and speed tells you whether the alert arrives while it still matters.

Foundation

Built around the NVIDIA stack.

Every idea maps to NVIDIA tools that already exist, so effort goes into the security problem instead of rebuilding plumbing.

Idea	NVIDIA tie
Private AI Appliance	NIM, NeMo Retriever, NeMo Guardrails
Firearm and aggression detection	Cosmos, Omniverse Replicator, TAO, TensorRT, DeepStream
Edge to cloud	Jetson, Metropolis, DGX
Network defense	Morpheus
Crime Locater	RAPIDS cuDF
Open source VLM demo	Open vocabulary detection with a VLM; VLM agent on DGX

New to these names? The Learn page explains each one in plain language.

Roadmap

Next: package, measure, deploy.

Package

One installable appliance software image.

Measure

Publish speed benchmarks for every appliance tier.

Generate

Build the synthetic drawn weapon training dataset.

Validate

Publish a benchmark whitepaper.

Deploy

Stand up the Jetson edge to DGX reference design.

Report

Ship the CMMC compliance report from the network defense agent.

Results will be posted here as each step is completed.