Ve-AP: a governed AI agent for a car

Ve-AP: A Governed AI Agent in a Real Car Ve-AP : Vehicle edge - Agentic Platform framework Ve-AP is a prototype that lets a person ask questions about a car in natural language while keeping the AI agent behind a defined access boundary. It runs across two computers: an NVIDIA Jetson Orin Nano Super handles the agent, voice, driver monitoring, and road view; a Raspberry Pi 5 runs the hardware layer and cabin display. The project explores a practical question: How can an AI agent interact with vehicle hardware while making its permissions and refusals visible? How it works A request does not go directly from the language model to a vehicle device. It passes through a capability gateway that checks whether the requested operation is permitted. Each decision—including refusals—is written to a hash-chained journal. The hardware layer, called MHS, describes devices and their capabilities in a checked registry. It keeps four concerns separate: What the device can do How the capability reaches the device Where the device is connected Whether the request is admitted and meets device-level limits MHS does not grant permission. A separate gate decides whether a request may proceed, and a device interlock can still refuse it. What the demo shows In the car, I asked the system questions by voice and saw the answer alongside the readings it used. When I asked about a route that was not available, it said it did not know rather than inventing one. The demo also shows a driver-monitoring view that displays derived measurements and draws a face outline from numbers, rather than showing a camera image. The road-view page displays a dash-camera image after blurring the faces and number plates its detector finds. If blurring fails, the system sends no image. What was real—and what was simulated The demonstration used real cameras, a microphone, speech, car audio, cabin display, and battery-powered computers. The vehicle signals shown on the display came from a CAN simulator. The car’s OBD-II port was not connected. The dashboard labels the simulated readings accordingly. This distinction matters: the demo shows the software running in a car with real sensors and displays, but it does not show that Ve-AP reads a real car’s signals correctly. Hardware NVIDIA Jetson Orin Nano Super, 8 GB Raspberry Pi 5 with official 7-inch touchscreen USB webcam with microphone Wi-Fi dash camera Separate 12 V battery and DC-AC converters Car head unit connected over Bluetooth audio USB to CAN Modules