Voice Intelligence Platform: AI Agent Builds Multi-System Integration With Zero Human Code

The client needed an AI system capable of joining live Microsoft Teams meetings over the public telephone network, transcribing the full conversation in real time, and contributing a single, highly targeted question at precisely the right moment — with zero tolerance for poor questions. The domain is Action Learning, where peer enterprise leaders work through complex, high-stakes decisions together; one bad question out of a thousand is a failure. Security and data locality were non-negotiable: everything had to run and be stored in Europe. Coordinating four live external systems (Twilio, Teams, Supabase, n8n) across each other's undocumented edge cases, while delivering the quality bar required, represented a multi-system integration challenge conventional development approaches would have taken multiple developer-weeks to solve.

Fountain City applied a "human as director, AI as builder" development pattern using Claude Code CLI as the AI coding agent. The human set architectural constraints (Twilio owns ears and mouth, n8n owns the brain — no WebSockets, no custom audio processing), directed three independently testable milestones, and reviewed output. The agent authored every n8n workflow as JSON via REST API, every SQL migration against Supabase over IPv6, every TwiML template, Nginx config, and Cloudflare WAF rule — without touching any UI. Six systems were integrated: Twilio Programmable Voice for PSTN dialing, Deepgram Nova-3 for real-time transcription, n8n (self-hosted, Docker) for orchestration, Supabase for transcript storage, Microsoft Teams Audio Conferencing as the meeting room, and AWS EC2 + Cloudflare for hosting. The agent resolved eight significant undocumented integration edge cases autonomously — including ghost transcription sessions from Twilio sidecar processes, undocumented n8n Supabase node field names, and Teams DTMF timing variations — typically within 2–3 retry cycles. Question quality was refined through 100+ human-validated synthetic test conversations before live deployment. All data, models, and systems are either European or locally hosted.

A fully operational voice intelligence platform was delivered that dials into live Microsoft Teams meetings via PSTN, transcribes conversations in real time, and injects a single AI-generated question at the right moment — with all audio, data, and models hosted entirely in Europe. The entire multi-system technical build was completed with zero lines of human-written code, at a total infrastructure and telephony cost of under $5 for the build-and-test cycle. Eight real-world undocumented integration edge cases were solved autonomously by the AI agent. The human director's time was spent entirely on architecture, constraints, and quality review — not implementation. The system produced comprehensive documentation as a natural byproduct of the agent's working method, and is now being tested in live calls ahead of the next development phase focused on scaling to multiple parallel sessions. The client noted: "You have demonstrated a deep and effective understanding of our needs and Action Learning principles."