A conference talk walkthrough of building 'Pip', a quirky AI pair programmer IntelliJ plugin with an attitude. The system uses a Spring Boot/Kotlin backend with a local Qwen 3 8B model, LangChain4j, and a Qdrant vector database. It implements an agentic workflow with categorization, judge/validation, and response agents. RAG is used to inject personality from 23,000 Slack messages of a real colleague into the LLM context, using dense, sparse, and multi-vector embeddings with reranking. MCP integrations enable meme generation, Spotify control, and Git branch creation. The talk also covers honest reflections on LLM limitations: non-determinism, prompt fragility, context rot, sustainability concerns, and MCP security risks around tool chaining.
Sort: