This is how I’ve been running my homelab lately: a self-hosted Mattermost workspace populated by AI agents that work across my homelab and GitLab instance. I post a bug/feature request, research topic, or business idea, and an agent will spin up and execute end to end, all the way to an MR, 75-page research doc, or a deployed landing page and a market-validation plan.

It uses a custom orchestration service over slightly modded Claude Code and Codex sessions, containerized for environment isolation. The orchestrator manages thread context and persistent memory (WIP), provides each agent/thread a durable Git repo, facilitates mid-session steering, and even bridges connections to friends’ and family’s agents.

In the first two weeks, I merged ~240 of its MRs across 51 different projects, comprising ~60k lines of changes. With frontier models and tight verification loops, the work is consistently good enough to be highly parallelized and asynchronous.

An extremely (over)engineered personal cloud running mostly out of a server rack in my home-office closet. It spans 18 hosts (bare metal + VMs). The hardware and network layer includes a ~100 TB Unraid NAS; an OPNsense router, managed switch, and wireless access points; and an RTX 4090 rig for local AI inference. The software stack includes Proxmox for VMs; NixOS for declarative host configuration; self-hosted GitLab and CI; a three-node k3s cluster running ~80 pods; Vault for secret management; centralized logging and metrics; custom agent orchestrators; and a growing pile of other custom software.

Nearly all of it is declared in code and managed through GitOps, making it legible to the agents running inside it and allowing them to continuously improve the lab with minimal human bottlenecking.

A web-based solver for iMessage/GamePigeon Word Hunt. It uses OpenCV to identify the grid and letters and traverses a trie’d English dictionary to find all possible words, as well as summary stats like total available points and word count (finding these was the original motivation for this project). The app is completely client-based; all CV and solving logic happens in the browser through vanilla JS.

(I don’t even have an iPhone—I had to test with screenshots I found on Google.)

A self-improving, voice-enabled, GPT-4-based assistant with multiple interfaces, written in Python. The design is very loosely inspired by the human brain, and with decent success—the latency is not great, but the bot makes fine conversation, follows instructions, and, upon my request, gave itself the ability to search Google, scrape my screen with OCR, and control my Spotify, among other things.

A website that uses the rules of species counterpoint to compose countermelodies to user-provided melodies. Built on a RESTful API written in Python with Flask and a simple vanilla JS/HTML/CSS web frontend. The site features a demo if you don’t happen to have a suitable cantus firmus MIDI file on hand.

A web app that uses an evolutionary algorithm to generate music, with humans in the loop as a fitness test via a voting bracket. The actual algorithm for generating and combining songs is pretty naive, however, and I’ve not observed convergence on a pleasant melody in my use. I’d like to revisit this idea but try something like interpolating between songs in a latent space, which should give better results.

A tool that generates distance-optimized (as opposed to just flattened) 2D projections of the stars. Written in Python and based on the principles of self-organizing maps. Includes error indicators at both the star and map level. Generalized to I/O of any dimension, but with few use cases for anything above 3D. This is the first project I did a write-up on.

The first iteration of my custom LLM assistant. Built with Python/Flask and a Discord bot interface. This was in the early days of ChatGPT, and the goal was getting it to correctly solve problems with multiple steps and/or specific, objectively right answers. This really boiled down to using Google and being able to solve basic-ish math word problems. I had moderate success—it could accurately answer questions about things like celebrity ages, caloric content, and movie showtimes, and could correctly answer questions like “how long could I survive on 100 cans of black beans?” about 30% of the time. This was near-instantly obsoleted by the release of gpt-3.5-turbo and ChatGPT plugins.