Stemma

Overview

Stemma is a Danish-speaking AI phone assistant for inbound calls. The problem we keep seeing in small and mid-sized Danish businesses is the same: calls come in outside opening hours, during busy periods, or when the person who can answer is already on another line, and those calls turn into lost bookings. Stemma picks up, speaks Danish, and handles the conversation end to end: scheduling appointments, taking bookings, and closing simple sales.

What I built

• Real-time voice pipeline. ASR → LLM → TTS, tuned for Danish and for the latency budget of a live phone conversation. Most of the engineering work is the seams between those stages, not the stages themselves.
• Telephony integration. Wired the pipeline up to telephony APIs and the backend systems that actually take a booking on the other side, so the assistant doesn’t just promise an outcome. It makes it happen.
• Pilot deployments. We’re running with real customers, measuring missed-call conversion against their previous baseline. That feedback loop, what the assistant does well and where it fumbles, is what we ship against every week.

What I’m learning

The interesting problem is rarely the model itself; it’s the failure modes around the edges. A voice agent that’s “right” 95% of the time still loses customer trust on the 5% if those failures aren’t graceful. Most of my time goes into making the system fail visibly rather than fail confidently.