VoiceOps: AI agents that complete property-management workflows

Live demo: voiceops-1ayi.onrender.com. It runs the offline scripted model with a mocked legacy system, and it is on Render's free plan, so the first load can take about a minute while it wakes up. Use Chrome for the voice mic.

VoiceOps is a working voice and messaging agent for a property manager. It handles resident and prospect conversations over phone, SMS, web chat and email, and completes real operational workflows against a legacy system of record: collections and payment plans, payment links, lease renewals, tour booking and lead follow-up.

The interesting part is everything around the model:

• Guardrails in code, not in the prompt. Identity verification, two-step confirmation that checks the user's own words, policy limits, idempotency, scope limits and human escalation.
• Legacy integration that survives a bad day. OAuth refresh, timeouts, retries with backoff, a circuit breaker, typed validation of a messy vendor API, and durable writes through an outbox.
• One conversation across channels. A call can continue over SMS without verifying again.
• Compliant outbound. Consent, quiet hours in the resident's time zone, frequency caps and dispute holds.
• Evals, tracing and metrics. 17 graded scenarios, task success, containment, escalation precision/recall, latency, cost per conversation, and a regression gate in CI.

flowchart LR
  U[Residents and prospects<br/>phone, SMS, web chat, email] <--> TW[Twilio / browser console]
  TW <--> API[VoiceOps agent platform<br/>FastAPI]
  API <--> LLM[Claude<br/>or offline scripted model]
  API --> PMS[(Legacy property system<br/>system of record)]
  API --> DB[(Store: cases, audit log,<br/>outbox, consent, traces)]
  API --> HQ[Human escalation queue]

It runs fully offline with no API keys. A deterministic scripted model stands in for the LLM, and the legacy PMS is mocked, so you can demo it anywhere. Set ANTHROPIC_API_KEY to run the same agent on Claude, and add Twilio credentials to take real phone calls.

The original single-file insurance call simulator (v1), with its screenshot and demo video, is in v1/.

---

Quickstart

pip install -r requirements.txt
python -m voiceops            # console at http://localhost:8000 (mock PMS on :8001)

Open the console in Chrome, pick Maria Lopez · Unit 4B, press Start conversation, click the mic and say "This is Maria Lopez in unit 4B. Can I split my balance into three payments?"

Run with	How
Offline (default)	nothing to configure; the scripted model drives the same tools and guardrails
Claude	export ANTHROPIC_API_KEY=... (optional: VOICEOPS_MODEL=claude-sonnet-5-5, VOICEOPS_EFFORT=low|medium)
Real phone + SMS	Point a Twilio number's voice webhook at POST {PUBLIC_BASE_URL}/twilio/voice and SMS at /twilio/sms; set TWILIO_ACCOUNT_SID, TWILIO_AUTH_TOKEN, TWILIO_FROM_NUMBER, PUBLIC_BASE_URL (e.g. an ngrok URL). Signatures are validated when the auth token is set.
Frozen clock for demos	VOICEOPS_NOW=2026-10-27T10:00:00-07:00 (quiet hours and dates become reproducible)

python -m pytest -q                                   # 40 unit tests
python -m voiceops.evals                              # 17 eval scenarios, offline
python -m voiceops.evals --provider claude --trials 3 # the real agent, pass^k over 3 trials
python -m voiceops.evals --compare evals/baseline.scripted.json   # regression gate (CI)

The interview walkthrough is in docs/DEMO.md. Design notes are in docs/ARCHITECTURE.md.

---

Deploy to Render

render.yaml is a Render Blueprint for one web service that runs both the agent platform and the mock legacy PMS. In Render, choose New > Blueprint, pick this repo, and fill in the secrets it asks for:

• DEMO_PASSWORD (recommended): puts the console and API behind basic auth, with any username. Twilio webhooks stay open and are checked by signature instead.
• ANTHROPIC_API_KEY (optional): without it the offline scripted model runs.
• PUBLIC_BASE_URL, TWILIO_ACCOUNT_SID, TWILIO_AUTH_TOKEN, TWILIO_FROM_NUMBER (optional): for real calls and texts. Set PUBLIC_BASE_URL to the service's https://….onrender.com URL, then point the Twilio number's voice webhook at /twilio/voice and its messaging webhook at /twilio/sms.

Demo data lives in SQLite on the instance's disk, so it resets on each deploy or restart. That suits a demo. The free plan also sleeps when idle, so open the URL a minute before you share your screen.

How a turn works

flowchart LR
  subgraph Channels
    V[Voice: browser mic / Twilio] --> R
    S[SMS: Twilio / simulator] --> R
    W[Web chat] --> R
    E[Email inbound] --> R
  end
  R[Resolve contact to open case<br/>shared across channels] --> P[Redact card numbers / SSNs]
  P --> H{Hard triggers<br/>emergency, opt-out, legal,<br/>dispute, human request}
  H -- yes --> X[Escalate / reply<br/>no model call]
  H -- no --> L[Model + tools loop<br/>budgets, traced]
  L --> T[Tool layer<br/>verification, scope, policy,<br/>confirmation check]
  T --> C[PMS client<br/>auth, retries, breaker,<br/>validation, outbox]
  C --> PMS[(Legacy PMS<br/>system of record)]
  T --> A[(Audit log<br/>hash-chained)]
  L --> G[Output guard<br/>no account data before verification]
  G --> O[Reply on the same channel]

What maps to what

What the role needs	Where it lives
Agents that complete workflows and write to a system of record with an audit trail	agent/tools.py (prepare/commit writes), store.py (hash-chained audit log)
Shared state across voice, SMS, web chat and email	agent/runtime.py conversation_for + contact index, history rebuilt per turn in agent/prompts.py
Compliant outbound for collections, abandoned applications, leads and renewals	campaigns.py
Guardrails for payments and tenant records (confirmation, idempotency, scope, escalation)	agent/guardrails.py, agent/tools.py, tests in tests/test_guardrails.py
Legacy API integration (auth, retries, timeouts, validation, durable writes)	integrations/pms_client.py against

Project layout

voiceops/
  agent/          runtime loop, tools, guardrails, prompts, retrieval, Claude + scripted models
  integrations/   legacy PMS client, Twilio helpers
  legacy_pms/     mock vendor API with fault injection (the system of record)
  campaigns.py    outbound workflows + compliance gate
  store.py        agent-side state: conversations, audit log, outbox, consent, traces
  server.py       channel webhooks + console API
  evals.py        eval runner and regression gate
config/           per-org policy and handbook
evals/            scenarios, baseline, reports
web/console.html  operator console (voice via Web Speech API)
tests/            unit tests, including a misbehaving-model suite
docs/             interview demo script and design notes
render.yaml       one-click Render deployment
v1/               the original single-file insurance call simulator

Security notes

• The agent never handles card data. Payments go through a secure link; card numbers, CVVs and SSNs are redacted before text reaches the model, logs, traces or storage (an eval checks every table).
• Caller ID and sender address identify a person but never verify them. Account tools require verify_identity, and an output guard blocks account figures in replies to unverified callers.
• Twilio webhooks are signature-checked when TWILIO_AUTH_TOKEN is set. The console API has no auth: it is a local demo tool and must not be exposed publicly as-is.