Paula Dozsa explains the engineering constraints behind a voice-first AI companion: short response budgets, stable character behaviour, and recognising when a user is interrupting rather than merely acknowledging the speaker. Paula Dozsa describes measuring latency stage by stage and routing emotionally consequential conversations to more capable models instead of treating routing solely as a cost optimisation.
Paula Dozsa outlines a memory system that retrieves relevant facts and preferences, compresses older material, and rebuilds context for each turn. Separating stable character information from changing conversation state helps preserve persona while allowing the system to incorporate new memories and app context.
Paula Dozsa also describes a development workflow with distinct implementation, review and pull-request supervision agents. Automated character evaluations and explicit feedback loops support that workflow, while human engineers still decompose tasks, define checkpoints and inspect results. Reported improvements are presented as the team's observations rather than independently established causal effects.
Watch on YouTube
