ProjectBEA

The story

ProjectBEA started as a simple experiment to learn Python properly and turned into one of the most complex thing I have ever built: a modular AI persona engine. It runs Bea, a single always-on consciousness that talks out loud, plays Minecraft on a vanilla server with other people, holds conversations on Discord, Telegram and Twitch, remembers who you are across sessions, and works through the objectives her owner sets for the stream. The architecture is the part I care about most.

Every sense pushes onto one perception bus. An attention gate decides, in pure rule code, whether something deserves a thought, a passing note, or nothing at all, so a busy chat costs almost nothing. What survives the gate reaches one mind that reasons and acts through tools, running one turn at a time per channel but several channels in parallel, so she can talk on stage and text at the same time.

Everything she says leaves through a single expression sink that drives TTS and the OBS avatar. Every capability beyond that is a plugin skill that can be toggled at runtime: Memory, a RAG layer with a diary, person cards and self-lore over one SQLite file with local embeddings; Minecraft, a fully autonomous agent that plays toward objectives through a custom Fabric mod I wrote that exposes the entire game state over WebSocket; Discord, a cross-language voice pipeline where a Node js bot captures live audio and streams the answer back with barge-in and resumption; plus Telegram, Twitch, Donations, Stream Plan, Social and Dream. LLM, TTS, STT and OBS all sit behind interfaces, models are configured per role in pools that round-robin and fall back when a provider dies, and almost everything hot-reloads without a restart.

On top there is a React and FastAPI control room: a bento overview, the attention gate drawn live, today's objectives, her memory, every ability and every setting. It ships with around 600 tests that run with no network and no API keys. This is the project that turned me from someone who writes scripts into someone who can architect a large, multi-module, multi-language system, and it is open-source and self-hostable.