Problem
Large enterprises struggle to surface accurate information from fragmented knowledge bases. Standard RAG pipelines hallucinate when queries require multi-hop reasoning across entities and relationships — a pattern common in enterprise contexts.
What I built
Architected and deployed a production-grade multi-agent system using CrewAI, LangGraph, and LangChain. The system routes queries through specialist agents — a graph traversal agent operating over AWS Neptune and a semantic search agent — whose outputs are synthesised into grounded responses.
Designed GraphRAG pipelines on AWS Bedrock and Neptune to improve retrieval precision. Built cloud-native orchestration infrastructure using AWS Lambda, Step Functions, and Terraform to handle long-running, distributed agent workflows reliably across environments.
Optimised API design and data access patterns across the system.
Impact
- 90% reduction in redundant LLM and API calls through optimised data access patterns
- Significantly lower latency and compute costs in production
- Improved retrieval precision and reduced hallucination rates in LLM-driven tasks
- Delivered from requirements discovery through pilot deployment to production rollout