ORBIS 2045 is a standalone hardware appliance that runs a local LLM and controls your phone, entirely offline. No cloud dependency, no API subscriptions, no data leaving your desk. It sits on your desk as a physical statement that intelligence doesn't need a data center.
The Stack
Compute
Raspberry Pi 5 (8GB) running a quantized 7B-parameter model via llama.cpp with GPU offload to the VideoCore VII. The 4-bit quantized model achieves ~12 tokens/second on the Pi 5 — not fast enough for real-time conversation but perfectly adequate for task-level inference where latency tolerance is 2-5 seconds. The VideoCore VII offloads approximately 40% of matrix operations, reducing CPU-bound inference time by 35% compared to CPU-only.
Networking
Mesh-network API for peer-to-peer inference routing. Multiple ORBIS units can form a self-healing mesh using Wi-Fi Direct and BLE neighbor discovery. When a device's local model is insufficient for a complex query, the request is routed to a peer with capacity, and the response is routed back. No external API keys required at any layer — the entire network is air-gapped from the public internet.
Mobile Control
ADB over Wi-Fi with a custom accessibility service that maps natural language commands to UI automations. The phone connects to the ORBIS over the local network, and the device issues ADB commands just like ORCA's execution layer, but entirely locally. The accessibility service runs as a persistent foreground app that reports screen state back to ORBIS for context-aware automation.
Interface
Neopixel LED strips embedded in the chassis pulse based on inference load. Blue means idle. Amber means chain-of-thought reasoning in progress. Red means heavy batch processing. The LEDs are driven by a small microcontroller (RP2040) that monitors the Pi's CPU/GPU utilization via I2C, ensuring the light patterns are responsive even when the main CPU is saturated.
Industrial Design
The chassis is 3D-printed with a matte black PETG finish and angled geometry that looks like it was pulled from a sci-fi prop department. The ventilation grilles are designed as parametric hexagonal lattices optimized for passive airflow. The front panel has a single amber LED strip that doubles as a status indicator and ambient light. It is a statement device. It sits on your desk and reminds you that intelligence doesn't need a data center.
Dimensions: 180mm x 120mm x 80mm. Weight: ~420g with the Pi and cooling fan. Power: 5V/3A via USB-C, consumes ~12W at peak inference load.