Airman VisionResearch
Benchmark what actually cuts false alarms.
Security AI makes a lot of claims. Every figure Airman Vision publishes will come from our own runs on named NVIDIA hardware, with the scripts published alongside.
False alarm study
Four approaches, measured on NVIDIA GPUs.
| YOLO family | Fast single pass detector. Measure recall, false alarms per day and latency. |
| CNN classifiers | Verify detector crops. Measure the precision gain over the detector alone. |
| Vision transformers | Global context on the clip. Measure hard case accuracy and GPU memory. |
| Multimodal AI (VLM) | Language guided review. Measure suppression rate, missed threat rate and cost per clip. |
Output: a published benchmark and the production false alarm filter. Results are reported as measured, including where a method loses.
Benchmark whitepaper plan
Speed numbers, with and without TensorRT.
| Workload / comparison | Hardware | Metrics |
|---|---|---|
| LLM serving Ollama versus NIM | RTX 5090, DGX Spark and RTX PRO 6000 | Tokens per second, time to first token, concurrent users |
| Detector PyTorch versus TensorRT | RTX 5090 system and DGX | Frames per second, latency, GPU memory |
| VLM second check | RTX PRO 6000 and DGX | Seconds per clip and cost per 1,000 alerts |
| Analytics pandas versus cuDF | RTX 5090 system and DGX | Speedup on one dataset |
Foundation
Every module maps to a named NVIDIA component.
| Private AI | NIM, TensorRT LLM, NeMo Retriever, NeMo Guardrails, Nemotron |
| Detection | Cosmos, Omniverse Replicator, TAO, TensorRT, Triton |
| Streaming and edge | DeepStream, Metropolis, Jetson |
| Search and summaries | Video search and summarization blueprint, NeMo Retriever |
| Cyber defense | Morpheus phishing detection and digital fingerprinting |
| Analytics | RAPIDS cuDF |
| Hardware | DGX Spark, RTX PRO 6000 Blackwell, DGX, Jetson |
New to these names? The Learn page explains each one in plain language.
Roadmap
Prototype. Benchmark. Pilot. Ship.
Prototype
Private AI appliance and local agent running. Detection pipeline stood up on synthetic data.
Benchmark
False alarm study, TensorRT and cuDF numbers, and a published whitepaper.
Pilot
Two to three design partners on Team systems running vision, private AI and the network agent.
Launch
General release across all three system sizes, the CMMC agent, and Jetson edge units.
Results will be posted here as each phase is completed.