How a single flag in llama.cpp turns a 35B Mixture-of-Experts model from OOM to 23 tok/s on a 12GB GPU.

Benchmarking Multi-Token Prediction (MTP) on Qwen 3.6 27B via llama.cpp on an RTX 3090 — 1.5× speedup in agentic tool-call chains.

A web dashboard for 96 PR reviews — built with Python stdlib, Chart.js, and zero external server dependencies.

After building an LLM-powered issue manager, the natural next step was PRs — syncing, diffing, reviewing, and posting feedback with the same local-first pattern.

A sequel to the issuer architecture post: how SQLite state, config.yaml, kanban workflow, and LLM-assisted judgment made large-scale issue triage safe.

issuer is a local Python CLI that uses SQLite and LLMs to triage GitHub issues — sync, analyze, review, close.

A Ghostty tab started printing raw escape sequences for every key, including mouse moves. Here is what causes it, why iTerm2 never showed it, and a six-line zshrc hook that heals it.

A direct RTX 3090 benchmark comparing my llama.cpp plus llama-swap Qwen3.6-27B setup against the club-3090 vLLM tools-text solution.

A vision-language-model-driven UI test typed my password into the username field. Fixing it took eye-icon prompts, before/after-click diffs, canary-byte typing, and uncovering two latent bugs.

How to fit NVIDIA's Nemotron-3-Nano-30B multimodal model with vision support on a single RTX 3090 — benchmarks, VRAM tricks, and the surprising resolution behavior.