All Posts

A self-hosted ChatGPT

I deployed Open WebUI ChatGPT-style chat UI for every device on the LAN, and a shared RAG-over-MCP retrieval server that both the chat and the coding agent can attach to.

Read More

A household coding agent

Running an open-weights coding agent entirely on the home LAN -- a spare box serves a Qwen3-Coder MoE over llama.cpp to laptops and phones around the house. Why the harness is Qwen-Code, and how a quantization sweep found the fastest model that still codes.

Read More

Bitcoin regime detection with HMMs

A 3-state Gaussian Hidden Markov Model identifies bear / ranging / bull regimes in BTC from a single causal feature -- drawdown from the running all-time high -- with median bear-posterior of 1.00 across calendar 2022 and exactly one Viterbi flip. Why drawdown beats log-return and log-price as the HMM observation.

Read More

Spread on TV

Spread has been featured in the Belgian news broadcast VRT NWS

Read More