Learn

The terms, in plain language.

AI and security come with a lot of acronyms. This page explains the ones used across Airman Vision without assuming a technical background.

Core ideas NVIDIA tools Open source tools Security terms

Core ideas

How the pieces think.

Local AI: AI that runs on a computer you own, in your building, instead of in a vendor's data center. Your questions and files never leave.
LLM (large language model): The kind of AI behind chat assistants. It reads text and writes text. On the appliance it runs locally.
Computer vision: AI that looks at images or video and recognizes what is in them, such as a person, a vehicle or a weapon.
VLM (vision language model): A model that understands both pictures and words. It can watch a clip and answer a question about it, which is why it makes a good second check.
Open vocabulary detection: Finding things from a plain description, like "red jacket near door 3", even if the model was never trained on that exact label.
Retrieval (RAG): Letting a language model look up the right passages in your documents before it answers, so replies are grounded in your files and can cite them.
AI agent: A model that can take actions, like clicking in a browser or opening a file, not just chat. At Airman Vision a person approves each action.
Synthetic data: Training images created by simulation rather than recorded by a real camera. Useful when the real event, like a drawn weapon, is rare.
False alarm rate: How often the system raises an alert when nothing is actually wrong. Lower is better, and it matters as much as accuracy.
Edge, site and core: The edge is the small computer at the camera that reacts instantly. The site server does the second check and search. The optional core reviews, learns and sends improved models back out.
Benchmark: A repeatable test that measures accuracy, false alarms and speed so different models and hardware can be compared fairly.
Inference: Running a trained model to get an answer. Training teaches the model; inference is the model doing its job.
Precision and recall: Precision is how many alerts were real. Recall is how many real events were caught. A good detector needs both.
Mean average precision (mAP): A single score that summarizes how accurately a detector finds and boxes objects across many test images.
Held out test set: Real footage kept aside and never used for training, so the score reflects how the model does on video it has never seen.
Mixture of experts (MoE): A large model built from many smaller expert parts, where only a few run for each word. Big capability, lower compute per answer.
4 bit quantization: Storing a model's numbers in fewer bits so a much larger model fits in the same GPU memory, with a small accuracy tradeoff.
Capture port: A network port that receives a mirrored copy of traffic so the system can watch the LAN without sitting in the path.
PII and jailbreak rails: Safety rules that stop a model from exposing personal information or being tricked into ignoring its instructions.

NVIDIA tools

The stack underneath.

NIM: Prepackaged, optimized containers for running AI models. Think of it as a ready to run engine for a given model.
NeMo Retriever: NVIDIA's toolkit for searching documents so a language model can find the right information before answering.
NeMo Guardrails: A way to write safety rules around a language model: what it may discuss, what it may do, and what it must refuse.
Nemotron: A family of NVIDIA language models, one of several used in the appliance prototype.
Cosmos: NVIDIA models that generate realistic video of the physical world, used here to create training scenes.
Omniverse Replicator: A 3D simulation tool for producing labeled synthetic images with control over lighting, angles and objects.
TAO: A toolkit for training and fine tuning vision models without building everything from scratch.
TensorRT: Software that optimizes a trained model so it runs much faster on NVIDIA GPUs.
DeepStream: A framework for running AI on many live video streams at once.
Metropolis: NVIDIA's platform and partner ecosystem for vision AI in cities, buildings and security.
Jetson: Small, low power NVIDIA computers built to run AI at the edge, next to the cameras.
DGX: NVIDIA's large AI systems used for training and heavy analysis at the core.
Morpheus: An NVIDIA framework for cybersecurity AI that inspects network data at high speed.
RAPIDS cuDF: Lets data analysis that normally runs on a CPU run on a GPU instead, often many times faster.
Video search and summarization blueprint: An NVIDIA reference design for searching and summarizing video with language, used for plain English footage search.
TensorRT LLM: The version of TensorRT tuned for large language models, used under NIM for fast private chat.
Triton Inference Server: NVIDIA software that serves many trained models at once and handles requests from cameras and apps.
DGX Spark: A compact NVIDIA desktop AI system with 128 GB of unified memory, the basis of the Core reference system.
RTX PRO 6000 Blackwell: A professional NVIDIA GPU with 96 GB of memory, used in the Team and Enterprise reference systems.
RTX 5090: A high end NVIDIA graphics card. The local agent prototype runs on one today.

Open source tools

Used in the prototypes.

Ollama: A simple way to download and run open language models on your own computer.
Open WebUI: A chat screen, similar to familiar AI assistants, that talks to local models.
Qwen and Gemma: Open weight language model families that can run entirely on local hardware.
MCP (Model Context Protocol): An open standard for connecting AI models to tools, such as a browser or files on a PC.
YOLO: "You Only Look Once", a popular family of fast object detectors for real time video.
CNNs and vision transformers: Two different designs for image models. Part of the research is measuring which works best for security video.

Security terms

Why it matters.

CMMC practice families: Groups of related security requirements, such as access control, audit and accountability, and incident response.
Evidence pack: A dated bundle showing what was monitored, what fired and what was done for each control, ready for an assessor.
CMMC: Cybersecurity Maturity Model Certification, the Department of Defense standard that suppliers handling defense information must meet.
Control mapping: Matching technical findings to the specific requirements, or controls, an assessor checks.
Phishing detection: Spotting fake emails or links designed to steal credentials or plant malware.
Digital fingerprinting: Learning the normal behavior of each user and device so unusual activity stands out.