Agentic AI Workflow @ Postman
End-to-end multi-agent system for semantic code retrieval
The Problem
Postman needed a way for AI agents to intelligently retrieve relevant code snippets from large codebases. Existing keyword-based search was too noisy, returning irrelevant results and slowing down agent workflows.
Approach
I designed a multi-agent architecture using CrewAI for orchestration and LangChain for pipeline composition. The system uses Tree-sitter for AST-style code parsing, breaking files into fine-grained chunks with metadata. These chunks are embedded using OpenAI's text-embedding-3-large and jina-embeddings-v2-base-code, then stored in ChromaDB for semantic retrieval. Agents decompose tasks, query the vector store, and reconcile results through feedback loops.
Results
- 40% reduction in agent query latency through optimized semantic retrieval
- 10,000+ fine-grained code embeddings indexed with sub-file granularity
- Automated CI/CD with GitHub Actions, reducing manual release effort by 2 hours per cycle