Case study · Nexxt Nest Group

SIYA, a governed agent system that is connected to a real phone line

Not a chatbot: a production voice AI plus ten orchestrated agents, built so nothing consequential happens without a human signing off.

Live in productionSaroj Sanjel, founder: architecture, code, security, operations (one developer)Call us · (03) 4517 8310 ↗

The problem

Small businesses lose jobs to missed calls, and most "AI agent" products either do too little (a chatbot with a script) or too much (agents acting autonomously with nobody accountable). SIYA had to solve both halves: a phone that always answers, and agents that can discover and prepare work but never act on money, contracts or deletion without human sign-off.

The agent system, from research to learning

SIYA is a loop, not a script. Oathline, our governed-agent framework, puts every agent through the same approval gate; Oathline Agents is open by waitlist only. Every pass through the loop leaves an auditable trail:

  1. Research: agents study the business, its channels, competitors and search presence
  2. Evidence: findings are recorded with sources, and claims without evidence don't advance
  3. Memory: leads, calls, outcomes and decisions persist across 94 database tables
  4. Tools: telephony, email drafting, CRM, calendars and web research, each tool-scoped per agent
  5. Decision: the orchestrator proposes the next best action as a typed intent
  6. Approval: intents touching money, contracts, sending or deletion queue for a human
  7. Action: approved actions execute; everything lands in the audit log
  8. Outcome & learning: results are tracked and feed the next recommendation

Governance: the authority model

Every agent intent is classified into an authority level. Reading and researching runs freely; drafting runs freely but output is held; anything that would contact someone, spend money, sign, delete or deploy is intercepted and queued for human approval. Email she writes is drafted, never auto-sent; the one automatic email is a fixed receipt to a new enquiry, where a client asks for it. Secrets are redacted from all agent output. The approvals ledger and audit log are first-class features, not afterthoughts.

governed agentspropose intentsAuthority checkclassifies L0–L4L0 · Read & researchruns freelyL1 · Draft contentruns freely, output heldL2 · Contact / sendqueued for approvalL3 · Spend / contractqueued for approvalL4 · Delete / deployqueued for approvalHumanapprovalapprove · edit · rejectEvery decision is written to an approvals ledger and audit log.

The hard engineering: voice at the protocol level

Twilio's ConversationRelay hands you a raw WebSocket and no framework. The server is an RFC 6455 WebSocket implementation built on Python's standard library, with zero third-party dependencies, including Twilio's HMAC-SHA1 request-signature validation. One gap needed closing: ConversationRelay's WebSocket carries no signature scheme of its own, so the TwiML response mints a single-use, 60-second relay token that the WebSocket handshake must present. On top of that: real-time barge-in (interrupt SIYA mid-sentence and she stops), DTMF handling, and latency tight enough for a live phone conversation.

Callerreal phone lineTwilioHMAC-signed webhookTwiML responsemints single-use tokenWebSockettoken checked at handshakeLLM + speechbarge-in interrupts mid-replyConversationRelay's WebSocket has no native signature scheme. A single-use,60-second relay token closes that gap. Implemented on the Python standard library, no frameworks.

Production operations

SIYA runs as a systemd service on a Linux server behind a Cloudflare Tunnel that whitelists a handful of API routes and 404s everything else at the edge, with a watchdog process, dated database snapshots and a documented backup/rollback procedure. Operational ledgers show real usage: 1,600+ processed messages, 1,000+ approval decisions recorded and 150 calls in the call log (1 Oct 2026). The codebase carries 360+ test files and 600+ commits since June 2026, including commits by a governed overnight agent whose work is reviewed under the same approval model as everything else.

What this project proves

Real-time voice AI built at the protocol level, a multi-agent system with genuine governance rather than vibes, and the result operated in production by one developer, end to end.