Skip to content

Notes

Experiments, written down.

Short records of what I tried and how it went. I leave the failures in.

rss feed

  1. August 2026

    Experimenting with local LLM inference

    One Qwen3.6 35B A3B serving both the assistant and bulk enrichment on two consumer GPUs: what fits, why the logs lie about which model ran, and the thinking-mode setting that returned empty responses.

  2. July 2026

    Improving hybrid retrieval for Globirie

    Where Globirie's lexical and dense retrieval actually stand, the pgvector filter trap, and why the reranking window matters more than candidate depth.