BoardBreeze Concierge

BoardBreeze Concierge is a 24/7 voice + SMS customer success agent for BoardBreeze (appboardbreeze.com), the AI meeting-minutes app I shipped earlier this year. Built in one week for the Anthropic × Cerebral Valley Global Hackathon, on Claude Opus 4.7. THE PROBLEM BoardBreeze grew faster than I could support it. Subscribers across the US, Canada, England, the Philippines, and Zimbabwe most of them non-technical board clerks, HOA managers, and nonprofit coordinators want to talk to a person on the phone. Some of their questions are easy. Some are even in the FAQ. But I am one person with a full-time civil-service job at City College of San Francisco. I cannot answer every call fast enough. Customer success at that scale needed a real solution. THE SOLUTION One phone number. One Claude Opus 4.7 agent. Five specialist modes inside that single agent, invoked by the agent based on what the caller needs: • Governance Helper — open meetings law, Brown Act, agenda posting rules • Product Expert — BoardBreeze features, pricing, plan comparisons • Tech Support — onboarding and account questions • Sales Closer — answers prospects, books demos • Escalation Handler — texts me directly when a human is genuinely needed That last mode is my dream come true. As an admin assistant, having a system that only bothers you when it truly matters? That's the innovation. Call it yourself: 1-844-786-2076 UNDER THE HOOD • Claude Opus 4.7 — voice and SMS conversation, via the Anthropic Messages API • Claude Haiku 4.5 — fast fact-checker inside the citation-verification tool. Opus thinks, Haiku verifies. • Direct Messages API on voice — I measured ~6 seconds of overhead on the higher-level agent stack, which is a hang-up on a phone call. So the voice channel calls the Messages API directly with sentence-by- sentence streaming. The caller hears the answer beginning while the answer is still being written. • Twilio — voice + SMS • ElevenLabs — text-to-speech in my own cloned voice • Supabase + pgvector + Voyage AI — governance knowledge base with vector search • Citation guardrail — every statute reference is verified to exist AND to support the claim before being read out loud. No hallucinated laws. • Cross-session memory — if Jane texts on Monday and calls on Thursday, the agent remembers her ARCHITECTURE: ONE MANAGED AGENT, TWO TRANSPORT LAYERS The Concierge is a single Claude Managed Agent (CMA) — one agent name, one system prompt, one set of registered custom tools. Following the "one agent + many skills" pattern Anthropic's Michael Cohen described at the hackathon Thursday session. • SMS runs entirely on CMA. Each caller's phone number has its own CMA session. Sessions persist across days and weeks, which is how the agent remembers a caller from a previous conversation — no custom memory layer required. CMA does it. • Voice reuses the same CMA agent definition (system prompt + tools + handlers) but executes the turn loop against the Anthropic Messages API directly, because I measured ~6 seconds of CMA event-stream overhead per turn — unworkable inside Twilio's 15-second webhook ceiling. • Four custom tools registered with the CMA agent: search_governance_kb, search_product_kb, verify_citation, escalate_to_grace. • The agent evolves in place via client.beta.agents.update() — added a tool mid-hackathon without resetting any existing caller's session memory. PRODUCTION DEPLOYMENT Pivoted from AWS App Runner to AWS ECS Fargate mid-hackathon. Docker containerization, Amazon ECR, AWS Secrets Manager for 11 API keys, two least-privilege IAM roles, Application Load Balancer with Amazon-issued TLS certificate, custom domain at concierge.appboardbreeze.com via Vercel DNS, CloudWatch logs, deployment circuit breaker with auto-rollback. Live at: https://concierge.appboardbreeze.com Built with Claude Code as my development partner.