Inspiration
Daniel Kahneman’s groundbreaking Thinking, Fast and Slow laid the foundation for understanding how humans think: through two systems — fast, intuitive System 1, and slow, analytical System 2. Today’s voice AIs largely mimic System 1: they’re responsive and fluid, but often shallow, sycophantic, and inaccurate. We were inspired to build an AI voice system that better mirrors how humans really think — combining instant, conversational interaction with deeper reflection and correction — so that even real-time AI can still be trustworthy, nuanced, and helpful.
What it does
(Fast): A voice agent that provides immediate, conversational responses — ideal for fluid, real-time interaction. This model mimics System 1 thinking: fast, intuitive, and responsive.
(Slow): A Claude reasoning model that processes the full transcript of the conversation asynchronously. It performs deep chain-of-thought reasoning to detect contradictions, hallucinations, or overlooked context — mimicking System 2 thinking. It interferes and comes in whenever needed.
How we built it
Voice - OpenAI real-time API Type 1 model: GPT 4o Type 2 model: Transcribe model (text-to-speech) -> Claude 3.7 Sonnet Reasoning model
Challenges we ran into
time issues, almost impossible to ship in 1.5 horus
Accomplishments that we're proud of
What we learned
Not all insights need to be voiced immediately. Sometimes delayed understanding is better for user trust.
What's next for Thinking Fast and Slow
In this future, we are looking to extend this via MCP. We aim to support tool and plugin integration for bigger contexts apart from voice such as calendars and other knowledge bases. We want context-aware voice agents that are NOT hallucinating.
Log in or sign up for Devpost to join the conversation.