Project Overview
Designed and shipped agentic AI systems for a science team: a LangGraph multi-agent architecture combining an internal
agent (SQL and report Q&A) and an external research agent (scientific APIs and websites), plus a production
document-extraction pipeline on AWS for multilingual medical reports.
Multi-Agent System (LangGraph)
Research Agent with Tool Calling
Document Extraction (OCR + VLMs)
LLM-as-a-Judge Evaluation
Key Responsibilities
-
Designed and built a full multi-agent system with LangGraph, end to end: orchestration (nodes, routing between agents, shared state), the agents and their tools, guardrails and evaluation.
-
Built an internal agent: a SQL agent and retrieval / Q&A over reports, with query validation and structured outputs.
-
Built an external research agent that answers scientific questions by calling research-paper and clinical-trial APIs and scientific websites (tool calling).
-
Ran experiments to evaluate agent behavior: automated LLM-as-a-judge evaluation comparing models and prompts on answer quality, latency and cost, with continuous improvement through new nodes and use cases.
-
Owned a production LLM extraction pipeline on AWS (OCR, translation, VLMs) that turns multilingual medical reports into structured entities, and built evaluation sets from real documents to benchmark and choose models.
Technologies
Python
LangGraph
LangChain
Multi-Agent Systems
Tool Calling
Structured Outputs
LLM-as-a-Judge
Guardrails
AWS
SQL
OCR
VLMs