Why msbel.com exists, and how a senior test automation engineer ended up running a Pi 5 in his apartment around the clock.
#notes#meta
Welcome to msbel.com.
This site is the umbrella for a few things I am building outside
of my day job. The current bench:
Alcyone — a self-hosted AI on a Raspberry Pi 5. Telegram in
and out, a paid model when it is awake on a real task, local
Qwen on Ollama for the heartbeat. The interesting part is what
it does between explicit requests.
Thalamus — a community plugin for OpenClaw that swaps
transcript-paste handoffs for a 3-field reference resolved
against a local vector store. MIT, on ClawHub and npm.
Trading bot — 24/7 on Binance Testnet with a $100 simulated
balance, never connected to real money. Nine signal layers, 46
ML features, daily CPCV retrain with a PBO gate.
QA Automation — an AI agent that authors Selenium and Gauge
specs from screen snapshots. The locator JSON is the single
source of truth.
Ember — an RPG written by hand in Unity after the
AI-scaffolded prototype proved unplayable.
The Pi runs Debian 13, hosts the trading bot, the agent stack,
and the encoder daemon. It restarts itself when load gets bad.
It is named Alcyone because that was the name I gave the AI
first; the box took the name later.
I will write here about what I learn from running this
continuously. The pieces I expect to publish first:
Why AI-generated games can’t be played, and what Ember
will do differently
What testnet ML trading actually taught me, the honest
version, including the parts where the model added no signal
The 9-layer signal stack, with the reasoning behind why
each layer earned its weight
Setting up a Pi 5 home server that can host a small SaaS,
with Cloudflare Tunnel, Caddy, and zero domain DNS pain
