RK.

Successor

Voice-based knowledge capture for semiconductor fab operations.

Problem

When an experienced fab employee leaves, years of troubleshooting knowledge leave with them. Nobody writes the PDF.

Approach

Captures knowledge through natural voice conversation and intelligent follow-up questions instead of documentation. A hybrid ranking algorithm on MongoDB Atlas fuses Atlas Search and Vector Search with Reciprocal Rank Fusion. The voice agent is a LangGraph state machine with ElevenLabs STT/TTS and a MongoDB checkpointer, so sessions are stateful and resume mid-conversation.

Results

Surfaces fixes from spoken symptoms across 60 fault classes and 185 fix outcomes. Built with Zubair Zafar and Swetha Prakash at MongoDB Build Fest, San Francisco.