PROJECT DOSSIER
2026
Echo Prismatic
Architect + engineer
A voice demuxer middleware. Derives speakers, emotions, tone, and more. Tracks conversation and enables dynamic agentic monitoring and intelligent tool calls.
voiceagentsai RustMQTTTauriSpeechmaticsHome Assistant
PROBLEM
A voice assistant that only transcribes words throws away most of what speech carries: who is speaking, their emotional state, their tone, and how the conversation moves over time. Without that signal an agent cannot reliably decide when to act, whom to trust, or which tool to call. The problem is turning a raw voice stream into live, structured signal that agents can monitor and act on.
APPROACH
Sit as middleware in the voice path and demultiplex the stream into derived channels (speaker, emotion, tone, and more) rather than collapsing it into a single transcript. Track the conversation as state and expose it so agents can monitor it dynamically and make intelligent tool calls against live context. Built adapter-bounded over an event bus, with resilience treated as first-class: every connection re-establishes its own state on reconnect.
EXECUTION
Built as a Rust workspace of adapter crates behind a core event-bus port. Each adapter derives a channel from the voice stream and publishes it onto the bus instead of collapsing everything into one transcript, so downstream agents subscribe only to the signal they need; speaker identification and auth gate who can do what. The conversation is tracked as state that agents monitor to decide when to act and which tool to call. The instructive arc was resilience: a broker blip once left the system connected-but-deaf, which generalized into a reconnect-and-re-establish pattern across every MQTT client, verified live by bouncing the broker under a running system.
OUTCOME
A voice demux middleware that turns a raw stream into live, structured signal (speaker, tone, conversation state) and lets agents monitor it and make intelligent tool calls, self-healing through outages.