June 2025 Summaries

Kwindla Hultman Kramer offers advice for developing voice AI agents, emphasizing the importance of understanding latency and instruction-following accuracy in the architecture. He suggests starting with a proven tech stack and progressively deploying to real users while optimizing cost and performance. Kramer highlights aiming for 800ms median voice-to-voice latency and advises on choosing the right models and tools, such as starting with models like GPT-4o or Gemini 2.5 Flash for effective function calling. The article discusses the necessity of using detailed prompts and context engineering to maintain instruction-following accuracy in multi-turn conversations. Kramer also underscores the importance of robust tooling for monitoring and debugging, the use of async function calls for efficiency, and the need for a reliable framework with good network transport, echo cancellation, and context management. He advises against starting with experimental models and stresses the importance of using widely accepted tools before delving into optimization.

Lemon Slice is building the next generation of video foundation models focused on humans, with a platform that allows anyone to create videos of expressive, talking characters. Their goal is to envision AI video not just as a creator tool, but as the future of interactive media and embodied AI. To achieve this, they developed a zero-shot model that supports 25fps streaming, which can generate real-time video of a character speaking with just a single image and audio input. Lemon Slice partnered with Daily to productize their model into a complete interactive experience, leveraging Pipecat's open-source framework for building multi-modal AI applications. The collaboration enabled Lemon Slice to focus on pushing the boundaries of AI video generation while handling complex infrastructure and architectural challenges.