Senior ML engineer obsessed with LLMs, evals and agents that run on hardware you actually own.
- LLM / GenAI systems and the evals that tell you whether they work (DSPy, GEPA, prompt optimization)
- Local-first coding agents: big models squeezed onto consumer GPUs, wired into real tools
- Self-hosted everything: home servers, pipelines, small tools that stay out of the cloud
- Games in Godot, when the brain wants a break
| Project | What it does |
|---|---|
| local-agent-kit | Run a 35B MoE coding agent on a 12 GB GPU. llama.cpp + llama-swap for opencode and Claude Code, with tuner and evals. Measured ~42 tok/s decode, 128k context on an RTX 3060, 10/12 on a Python eval suite. |
| redteam-lab | Local pentest lab for TryHackMe: LLM agent with guardrails, planner + worker orchestrator, web dashboard. |
| DeepMosaicsPlus | Remove mosaics from images and video on any GPU, with DirectML support for AMD. |
Also: llama.cpp, DSPy, systemd, Caddy.
Site: psychoastronaut.xyz
If something here helped you, you can buy me a coffee:
