High-school student. I build local AI infrastructure and the automation around it.
Most of what's here started as "I want my machine to do this by itself" and ended as a working system: local model routing, end-to-end generation pipelines that audit their own output, and a lot of Windows plumbing that nobody else wanted to touch.
Local-first AI stack — agent orchestration, model routing, and generation pipelines that check their own work.
| Repo | What it is |
|---|---|
| screendim | C# daemon: turn the physical backlight off via DDC/CI while screen capture & OCR keep working |
| local-ensemble-shim | OpenAI-compatible shim that ensembles two local models (multi-view + judge) with guards for real local-model failure modes |
| windows-pitfalls | Battle-tested PowerShell/VBS fixes: icon cache, autostart chaos, silent launchers, OEM codepage traps |
| condense | Compress long docs with a local LLM before feeding a cloud model — fewer input tokens, same answers |
measure it → break it → find the root cause → leave a handoff doc the next agent can use
- I orchestrate AI agents to ship systems — my unit of work is a system, not a function
- Everything gets a rollback path: backup → patch → verify → document what will silently break next time
- I care about the boring numbers: tokens, VRAM, watts, startup time, how many seconds before it degrades
Stack — PowerShell · Python · Node.js · C# · Ollama · llama.cpp · ComfyUI · SillyTavern
Currently: high school + too many background processes.