Vengram
About this project
In one sentence, what are you building?
Editing open weights to help a voice agent become self-evolving
What problem are you solving?
Self-evolving voice agent on Vapi that learns from calls by directly editing the weights of its underlying open-weight language model. An editor model reviews transcripts and tool calls, identifies recurring mistakes, and proposes targeted weight changes. We’re focusing on complex tasks like dermatology scheduling, where a single request can involve multiple providers, appointment types, and procedure-specific rules. Each proposed edit is evaluated on new cases and cases the agent previously handled correctly before deployment. Open source github attached that can be self-hosted with better models via Docker.