AmtHeld — Voice AI for Public Administration

AmtHeld is an AI-driven telephone assistant for German public administration: it answers citizen calls 24/7, detects the caller's language, and handles routine inquiries about documents, appointments, and services in more than 25 languages — with real-time transcription and GDPR-compliant handling on EU servers. I contracted on two fronts. On the voice side, I worked on the real-time speech pipeline that makes the assistant feel like a competent phone call rather than a menu tree — language handling, low-latency turn-taking, and conversation flow built on modern speech models. On the product side, I worked on the analytics dashboard that lets non-technical administrators see call volumes, topics, and transcripts at a glance. Government callers are the hardest audience a voice agent can face: every age group, every accent, high stakes, zero patience for lag. Building for that bar — inside strict German public-sector privacy constraints — shaped every technical choice. The work is under NDA, so this case study stays at the architecture level — I'm happy to go deeper in a call, within what the NDA allows.