CTBTO
Conversational AI for nuclear treaty diplomacy
Built a conversational AI prototype navigating Nuclear Treaty topics for the Comprehensive Nuclear-Test-Ban Treaty Organization.
A serious institution and a serious build. The CTBTO is the international body charged with ending nuclear weapons testing — not the kind of organisation that plays with conversational AI for the optics of it. Demonstrated at the Science and Technology Conference at the Hofburg Palace in Vienna: an anthropomorphic conversational interface holding the line on Nuclear Treaty topics in a diplomatic register.
The work tested whether such an interface can stay on-policy across the failure modes that sink most conversational AI in production: oversharing, off-script answers, role drift, hallucinated authority. Tavus for embodiment, Weaviate for retrieval, an open-weight reasoning model for the policy core, and a generative UI layer that adapts to the speaker.
The kind of build that informs how we train: where the cost of being subtly wrong is high, and the operating habits — review, scoping, governance — matter more than raw model capability.
// Want this kind of work inside your team?