Earshot
A WebRTC voice capture agent with typed tools and an interface that shows disagreement between speech recognition and model-written customer names.
Problem
Turning a conversation into structured records introduces two sources of uncertainty: what the recognizer heard and what the model decided to write.
Approach
Built a realtime voice interface with seven typed tools, closed-vocabulary validation and structured errors that guide the model’s next call. The public MVP visualizes tool actions and offers two repair paths when the recognized name and tool argument disagree.
Impact
Name disagreement appeared in six of seven manually observed sessions. That small observation motivated a visible correction interaction; it is not an accuracy benchmark or a measured reduction in CRM work. The expanded 2.0 implementation is user-confirmed but remains outside the linked public main branch.
Key Metrics
Technologies
Links
My Role
Designed and built the voice/tool interaction with AI assistance and wrote the phased 2.0 specification. Collaboration and handoff are recorded in the public repo; the user-confirmed 2.0 implementation is unpublished.