Currently working on.

Migrated this site from Next.js to Astro and gave it the terminal aesthetic it always wanted. Zero client JS on most pages, JetBrains Mono throughout, projects and case studies authored as MDX. The old version looked like every other student portfolio. This one looks like the kind of work I actually do.

Took over from OpenClaw as the agent harness I actually use day-to-day. Still running on a local model — issue-tracker triage and a few other quiet, repetitive jobs that don’t need a frontier API. The point isn’t to outperform anything; it’s to do the boring, frequent work without paying per token. Notes accumulate as I find the edges.

Benchmarking Qwen 3.5, Qwen 3.6, and Gemma 4 on the kind of internal tasks we actually want local inference for at Dewaweb — retrieval-grounded answering, structured extraction, short code edits. Public benchmarks are mostly noise for these use cases; what matters is throughput on our actual prompts and how often each model holds a system prompt. Writing it up as a decision memo, not a leaderboard.

Running my own hands-on evaluation of the agentic-CLI tools (Claude Code, Codex, OpenCode, Gemini CLI) and the current crop of local models (Qwen 3.5 / 3.6, Gemma 4 via Ollama). Less a benchmark, more a decision memo for myself — where each one excels, where they hallucinate, where the tool-use surfaces leak complexity, and which local models are actually realistic for everyday workloads. Notes keep getting reused as the tools update.

Ran OpenClaw locally for about two months as a thin agent harness on top of a local model — GitLab issue drafts and a handful of other small chores. It was a useful starting point but I’ve since moved over to Hermes (see entry above) as the daily driver. Keeping this here as a marker of the trajectory rather than as something active.