Tolan: Voice-First AI Companion — Paula Dozsa, Tolan

Building a voice-first AI companion requires prioritizing low latency and managing conversational volatility, as users frequently interrupt and switch topics. Maintaining a sub-two-second response time is critical for immersion, necessitating a tiered model architecture that routes high-stakes interactions to frontier models while using faster, cheaper models for casual turns. Effective memory management relies on a vector database that extracts and compresses facts, preferences, and emotional signals rather than maintaining a full transcript. Beyond the product itself, the team leverages a fleet of autonomous agents to automate code generation, peer review, and bug triage. This agentic workflow, which includes rigorous evaluation against production logs, has improved the app's crash-free rate to 99.9% and enabled rapid character development, demonstrating that managing AI agents requires many of the same organizational skills as managing human teams.