I build production AI systems: multi-agent workspaces, governed tool access, and the evaluation around them. Eighteen years shipping platforms in healthcare, insurance, and government. Right now that looks like Rivulets, the rivulet-dispatch library it is built on, a local Kokoro TTS wrapper, and daily writing at Signal Over Noise.
Rivulets is a local-first, peer-to-peer workspace in the shape of Slack. You put humans and teams of AI agents in the same channels. Dispatch is deterministic first, LLM only when a rule misses, with handoffs and loop guards so the roster cannot talk itself in circles. It sits on Agno AgentOS. There is no hosted service. Keys stay on the machine.
rivulet-dispatch is that routing core as a standalone MIT library. Mentions win. Deterministic rules run next. An injected LLM fallback runs only when nothing matched. A one-specialist lock and in-memory loop guards stop unsolicited rematch. Zero runtime dependencies.
kokoro-tts-api is a FastAPI wrapper around Kokoro-82M. Text in, WAV out. Voices are allowlisted. Weights stay on disk after the first Hugging Face download. CI boots the app without pulling the model.
Signal Over Noise is the public notebook: production AI, agent platforms, and what actually ships versus what gets announced.
| Repo | What it proves |
|---|---|
| Rivulets | Local-first multi-agent workspace, dispatch, MCP, P2P sync |
rivulet-dispatch  |
Testable multi-agent router. Mentions, rules, lock, loop guards. |
kokoro-tts-api  |
Local inference behind an HTTP contract. No weights in git. |
Open upstream until it merges: agno-agi/agno#9628 (team member messages in AgentOS session chat_history).
More public extracts are coming: a small MCP policy gateway, and system-level agent evals. The work is the record. The articles point at the repos.
- Writing: Christian Leadership
- Writing: Signal Over Noise
- Writing: The Vethrak Requiem
- LinkedIn: justinwilson411