We Built AI That Predicts Code Disasters (UC Berkeley Hackathon)

In this video, I showcase DiffSense - an AI-powered code analysis platform that predicts breaking changes before they happen, built at the UC Berkeley AI hackathon. I walk through everything from concept to implementation, including: Live demos: • Breaking change detection on real repositories • Intelligent code analysis with multi-expert AI • Real-time chat interface for repository exploration • Semantic search across complex codebases Full system overview: • Advanced RAG system with 768-dimensional code embeddings • Multi-expert AI personas (Security, Architecture, Performance, Quality) • Breaking change detection using AST parsing + ML prediction • Smart context curation (analyzes 200+ files, curates top 15 for AI) • Claude AI integration for natural language explanations This project combines cutting-edge AI with practical developer needs—semantic code understanding, predictive analytics, and intelligent automation to prevent production disasters before they happen. Built in 24 hours at UC Berkeley's AI hackathon, DiffSense represents the future of predictable software deployment. GitHub: https://github.com/jalenfran/DiffSense Website: https://www.jalencode.com/ LinkedIn: https://www.linkedin.com/in/jalen-francis/ #AI #MachineLearning #RAG #CodeAnalysis #Hackathon #SoftwareDevelopment #BreakingChanges #SemanticSearch #Claude #OpenAI #UCBerkeley Chapters: 0:00 Intro 1:02 What it does 2:09 How we built it 3:22 How to run 4:31 Demo Start 5:29 Running Analysis 7:47 Chatbot 11:40 Challenges 14:19 Accomplishments 15:26 What we learned 16:05 What's next 16:47 Conclusion