the underground · field notes
notes from a self-hosted fleet.
Honest writeups from running real AI on my own budget silicon — local LLM inference on Intel Arc, autonomous agents with guardrails that hold, and a memory brain that remembers across every tool. Real numbers, no hype, $0 cloud.
latest
Building a shared memory brain for AI agents
Every agent is an amnesiac — close the session and the context is gone. The fix isn't a bigger context window; it's a persistent, shared, self-correcting memory every tool reads from and writes to. The design decisions that make it work: supersede-never-delete, write the mechanism not the answer, and automatic write-back. Past 130,000 memories, live.
read the writeup →
Running 30B LLMs on a $500 Intel Arc B60
Serious local inference doesn't need a 4090. Here's the exact stack — llama.cpp, Vulkan, IPEX-LLM — the driver and VRAM gotchas that actually bite, quantization tradeoffs, and measured tok/s for a dense 30B-class model at every common quant. All at $0 cloud spend.
read the writeup →
more field notes in progress — the memory brain internals, guardrails that actually hold, and agent cost telemetry. subscribe via rss ↗