Getting Vocalinux — a fully offline, GPU-accelerated voice dictation tool — working on Debian, past three undocumented build hurdles Ubuntu users never see.
Read Article →
Running two purpose-tuned Ollama instances on one GPU — one for text, one for vision — fronted by Open WebUI as an authenticated API gateway, so neither Ollama ever touches the internet directly.
Read Article →
Upgrading Git-Kepo from an unreliable prototype to a production-grade NemoClaw agent: fixing Ollama context limits, splitting TOOLS.md to save tokens, and adding repo labelling and batch mode.
Read Article →
Building Git-Kepo, a Singlish-speaking NemoClaw specialist agent that clones repositories, runs churn/bus-factor/velocity diagnostics, and audits git history for leaked credentials using the LLM's own judgment.
Read Article →
Giving self-hosted OpenClaw agents real personality — building two Singlish-speaking AI characters on separate NemoClaw sandboxes that share a Matrix room and pass conversations to each other by mention.
Read Article →
Combining OpenClaw, NemoClaw, Ollama, and a self-hosted Matrix server into a fully private, kernel-sandboxed agentic AI stack with no third-party servers in the critical path.
Read Article →
Building a custom two-stage Dockerfile to run llama.cpp with CUDA on the NVIDIA Spark (ARM64/Blackwell), since the official images only target x86_64, plus multimodal model routing with a models.ini catalog.
Read Article →