Projects
What I've been building. Some are active, some are paused, a couple are old enough now to be archaeology. The blog at coffee-anon.com is where I write about them as they go. This page is the slower companion.
Hermes
An assistant that lives in the cluster and runs on a schedule. The project is partly an answer to a question I keep coming back to: how much can I do with local agents that I can fully trust with personal data? Hermes calls tools and sends messages to Signal, talks to the other agent on the cluster through an A2A surface I'm still figuring out the shape of, and writes anything that needs a human into an outbox. Most of the architecture is about keeping personal data inside the house.
A2A↔MCP bridge
Two protocols for getting agents to do things showed up at about the same time: Google's A2A for agent-to-agent communication and Anthropic's MCP for agent-to-tool. They cover overlapping ground but don't quite meet. The bridge sits in between, holds conversation state across both surfaces, and pauses for human approval before any destructive action. The reference implementation is public. The design notes are the blog's current focus.
llama-pod
Running large language models on consumer hardware, mostly to figure out what's actually doable at home now that the open-source LLM stack has caught up. llama-pod is the Kubernetes pod that does it: llama.cpp on a Strix Halo machine with multi-token-prediction speculative decoding. LiteLLM fronts the inference and routes requests from a mix of clients: Open WebUI and OpenCode for human use, plus a growing set of agents that lint commits, review CVEs, and otherwise quietly handle work in the background (Hermes among them). The hardware specs that work and the MTP gotchas are the durable learnings, and both will probably get their own pages here once the writeups firm up.
Scrapyard cluster
A Kubernetes cluster running on a rack of laptops and mini-PCs at home. It started in 2023 as one machine (jukebox, still around as the storage host) and grew out from there when one project needed to support many users connecting to a game server at once. A single host was the wrong shape for that. Most of the other projects on this page live on the cluster now. This page used to have more operational detail; I've trimmed it back because the world doesn't need another homelab runbook.
This Isn't Science
The blog at coffee-anon.com. Long-form writing on infrastructure and practice, from running systems to leading teams. The name is a reminder to myself: build and do things instead of over-analyzing them. Too much certainty too early, whether in a project or in a plan, misses the opportunities that only show up once you start doing the work.
Cartographer
An on-screen game companion, built as a web app with a server in the cluster. Designing for many simultaneous game-server connections made a single host (jukebox) feel too small for the workshop. The project is paused. The cluster that grew out of it is the more durable artifact.
Pipeline
A Dagster data pipeline for modeling prices in the EVE Online market. EVE has a thin-market real-time economy with a public API, which made it a good case for working through the mechanics end-to-end: scraping, transforming, columnar storage, dashboards. Paused for now. The git history is intact, and a writeup might come later.
Flask boilerplate
A starter project structure for Python/Flask web apps. I used it to bootstrap a handful of small projects in 2024 and 2025, then drifted toward other patterns. Public on GitHub, not actively maintained.
danjacobsen.com
My professional landing site at danjacobsen.com. One page, externally hosted, kept simple on purpose. Shares the palette with this site and the blog.