SELF-HOSTED AI / ENGINEERING LAB

Models under
my own control.

Exploring local inference, open-weight language models, model behavior, AI-assisted development, and the practical limits of self-hosted hardware.

Launch Drabsnarg AI ↗Authentication required

Local inference infrastructure

A 2017 iMac Pro with a 10-core Xeon W, 128 GB RAM, and Radeon Pro Vega 64 (16 GB VRAM) provides the foundation for experimentation with llama.cpp, Open WebUI, and Qwen coding models.

Testing includes CPU and Metal inference. GPU acceleration has not consistently improved performance on this hardware. No benchmark numbers are published without recorded measurements.

SYSTEM ARCHITECTURE

Authorized user↓Cloudflare Access↓Cloudflare Tunnel↓Open WebUI↓llama.cpp / Local model

Planned secure remote-access architecture. Public access and deployment state are not implied.

Research areas

Inference & Hardware

CPU, Metal, quantization, memory and throughput constraints.

Models & Evaluation

Model selection, context limits, task quality, and repeatable tests.

Coding Workflows