Voice agents with Realtime Video — Sidney Primas, LemonSlice
AI Engineer World's Fair 2026 · 26:36
Interactive AI avatars and real-time video generation
LemonSlice builds interactive AI avatars that add a visual layer to chatbots and voice agents. Developers can turn a single image into a talking, listening character, connect their chosen language model, voice provider and knowledge base, and deploy through an API or website widget. Applications include tutoring, customer support, entertainment and physical installations. LemonSlice Lite targets consumer avatar applications, while LemonSlice Pro provides higher-resolution characters for large-format displays.
Founded in 2024 by CEO Lina Colucci, CTO Sidney Primas and Andrew Weitz, LemonSlice develops its own video-generation models. Its LemonSlice-2 research introduced a causal video diffusion transformer and inference framework that animate human or stylized characters from an image and audio stream, including full-body gestures and surrounding scene motion. Causal attention, distillation-inspired training and caching support real-time generation; the company reported 20 frames per second on a single GPU for its 20-billion-parameter model.
LemonSlice raised a $10.5 million seed round in 2025, backed by investors including Matrix and Y Combinator. In a 2026 case study, the company reported that Y Combinator’s Office Hours Simulator scaled to hundreds of daily conversations across web and physical installations. Its Agora integration also supports Live Play Mobile’s 24/7 avatar gaming hosts.
AI Engineer World's Fair 2026 · 26:36
Affiliations reflect their AIE appearances, not necessarily current employment.
Primas described an API-based visual layer that connects to independently supplied language and voice models, with model orchestration as part of the discussion.
Affiliations reflect each recorded session, not necessarily current employment.