Trusting AI to Modernize Software at Scale
summary
Moderne builds a mass-scale refactoring pipeline on OpenRewrite and experiments with GPT-4o for recipe generation but finds raw LLM outputs often fail to compile, hallucinate methods, and produce non-deterministic edits. The successful stack uses retrieval-augmented embeddings with re-ranking, clustering-based sampling to recommend fixes across hundreds of thousands of methods, and a tool-calling assistant (via MCP/tool-calling) that runs, validates and summarizes recipe runs for large enterprise fleets; outcome: reliable search and recommendations, but generative recipe authoring still needs guardrails.
AI-written summary · about
chapters (13)
- 0:00Trusting AI for modernization
- 6:18Reliable code changes at scale
- 10:54Generative AI for recipes
- 16:53Variable reward limits in AI changes
- 18:18Improving search with RAG
- 23:16How AI search works
- 26:06Code recommendations from snippets
- 29:36Vendor usage in codebases
- 33:36AI assistant for recipe discovery
- 38:53Live demo of AI assistant
- 42:07Customer-controlled AI models
- 43:07Generating recipes from guides
- 44:10Balancing AI promise and reliability
related talks
34mModerne + OpenRewrite at Morgan Stanley: 5 Lessons From a Year at Scale
37mDeterministic modernization in an age of AI
40mHow BNY Uses Moderne to Make AI Agents Cheaper and More Accurate
54mHybrid Modernization: Combining OpenRewrite's Precision with LLM Intelligence for Spring
47mHow to Fix Your Technical Debt Dilemma
36mUnleashing Generative AI: Improving Developer Productivity in SRE
51mJava Meets AI: Empowering Developers to Build Intelligent Apps
43mAI w/o Spaghetti: Clean Architecture in the Age of AI