SYS-05·2026· CLASSIFIED
Echo
Multi-tenant voice-agent engine
“Speaks only what it can prove.”
A voice-agent engine that can be embedded on any website. A team of specialist agents answers only from verbatim, approved knowledge, and refuses safely when it can't.
- Agents
- 5
- Grounding
- closed-world
- Embed
- 1 script tag
- Speech
- local
Next.jsTypeScriptLocal router modelWhisperKokoroPiperEvals
CLASSIFIED · built for a client. Shown here as architecture and outcomes only: no names, no code.
The problem
A voice agent on a privacy-sensitive product can’t improvise. Every sentence it says has to trace back to approved facts.
What I built
- A crew of five specialists: Guide, Consent, Navigator, Finder and Handoff. A small local router model picks the specialist for each request, with a lexical fallback.
- Closed-world grounding. Answers are spoken only from verbatim, approved knowledge-base entries. If it isn’t in the knowledge base, the agent doesn’t say it.
- A consent classifier that fails safe. It checks quotes against the source, and when it’s unsure, the answer is no.
- Multi-tenant from day one. One engine serves many sites. A
/new-tenantagent skill sets up the next one. - One script tag embeds it on any website. A local speech server runs Whisper, Kokoro and Piper.
- An eval script per agent, so a change that makes an agent worse is caught before it ships.