Synapse - AI Knowledge System
Self-learning knowledge system connecting employees with internal experts using vector search and semantic retrieval
Overview
At TUM.ai Hackathon, I architected the backend for Synapse-a self-learning organizational knowledge system that transforms hidden company expertise into actionable solutions. Built in 36 hours with a 5-person team, I implemented a vector embeddings pipeline using OpenAI and pgvector for semantic search, an expert ranking algorithm with relevance and freshness scoring, and integrated Claude Sonnet for tool-calling chat functionality. The system enables employees to find internal experts who previously solved identical problems through natural language queries. We placed 2nd out of all teams, delivering a fully functional MVP with live demo.
Challenges & Solutions
- 1Building vector embeddings pipeline for semantic knowledge search in 36 hours
- 2Implementing expert ranking algorithm balancing relevance and freshness scores
- 3Designing semantic chunking strategy for optimal knowledge retrieval
- 4Integrating Claude Sonnet for natural language tool-calling interface
- 5Creating real-time expert matching system with PostgreSQL and pgvector
- 6Coordinating backend/frontend integration across 5-person team under time pressure
Results & Impact
Placed 2nd in TUM.ai Hackathon with fully functional MVP (not just prototype)
Built working semantic search engine connecting employees with internal experts
Delivered live demo with chat, voice, and video interview capabilities
Implemented scalable vector search architecture using pgvector and OpenAI
Created expert ranking system with relevance and temporal decay scoring
Shipped production-ready API with Prisma ORM and Express in 36 hours
More Views
