Terminal
NEXUS
A local AI lab that routes between three models, fully offline, on a laptop's integrated GPU.
- When
- Apr 2026 to now
- Built with
- Python
- FastAPI
- React
- Ollama
- LM Studio
- Vulkan
- models routed
- 3
- models routed
- models tested
- 15
- models tested
- tokens/sec on an iGPU
- 18-25
- tokens/sec on an iGPU
- cloud calls
- 0
- cloud calls
What I did
- Routes each request to Llama 3.1 8B, Llama 3.2 3B, or DeepSeek R1 by task type, and sends low-confidence answers back through a verification pass.
- FastAPI orchestration backend plus a six-view React dashboard with conversation memory across turns.
- Tested 15 candidate models, tracked down a Qwen3/Vulkan GPU bug, and tuned Q4_K_M quantized models for the Radeon 780M.