Graphene
Fluid UI and inference runtime for edge devices
Instead of static screens, Graphene lets an LLM compose a full native interface tailored to what the user actually needs. The model discovers what the device can do, builds the UI piece by piece, and adapts it on the fly. Everything renders and executes on the device itself.
Source: github.com/me-im-counting/graphene (private — contact for invite)
LLM discovers operators, builds a dashboard from "build me a depth estimation tool", then adapts it by adding a file browser on request.
LLM creates a reaction game with moving targets. Registers IPC triggers on button clicks, reacts to events, updates score in real time.
Three different apps — System Monitor, Color Picker, Task Manager — built back-to-back on the same runtime from different user prompts.
No templates
The model picks from buttons, lists, sliders, image views — whatever makes sense — and assembles the interface itself. Every app looks different because every request is different.
Closed loop
The model can query what operators exist on the device, read back any value it set, and register triggers that fire when the user interacts. It works with what's actually there.
Runs locally
The UI renders and inference runs on the device itself. DirectX 11, OpenCL, CPU. No server needed for the interface or the compute.
