A self-hosted replacement for kubesearch.dev that indexes more repositories, from more forges, with more manifest scanners — exposing the index through an HTTP API and an MCP server so agents can research how real homelabs configure their Helm charts. Written in Elixir on SQLite, so it runs comfortably on a free or cheap VM.

An extraction model post-trained from Qwen3.5-9B that answers a whole schema of typed fields in one parallel forward pass instead of generating tokens. Answers come back as read spans with exact character offsets, computed values, or calibrated distributions — so it can't re-spell or invent what isn't in the document.

A 4-node cluster running at home — three mini PCs (i5-9500T, 32GB RAM each) and a 2U rack server (24 cores, 384GB RAM) on Talos Linux and Kubernetes, fully git-managed with Renovate for automated image updates. A separate NAS on Proxmox manages 50TB of storage, with automated backups on intervals (the longest being 2 hours), kept as both local and remote copies.

Running and testing different LLMs locally, evaluating them across a range of tasks — building on the same muscle I use professionally for model and tool selection at Recital.

An AI-powered document classifier that automatically organizes documents and images. It extracts structured data from documents and makes them searchable — so finding a receipt or warranty info is a query, not a hunt. Designed to work with either a local LLM on your machine or a remote LLM, letting you choose your own tradeoff between privacy and accuracy.