Agentic Workflow Engine
Product · AI · LangGraph · Python
The compensation cycle is a multi-week, multi-step enterprise workflow that has to be executed carefully and auditably. Working with stakeholders, we framed the requirements and decided where a human must stay in the loop; I then built a LangGraph assistant that guides the cycle end-to-end without ever acting on its own at business-critical steps.
Problem
Running the compensation cycle meant coordinating many sequential phases with strict correctness and audit requirements — slow, manual, and easy to get wrong.
Solution
A LangGraph supervisor orchestrates 14+ phase agents with human-in-the-loop gates at business-critical steps, SSE streaming for live progress, shared front/back contracts, and Langfuse observability.
My Contribution
With the stakeholders, I elicited the requirements and defined which steps required mandatory human approval, then designed the agent graph and routing and built it — contributing as product owner and engineer.
Outcomes
- Reduced cycle time from weeks to hours
- Zero autonomous actions on business-critical steps
- Real-time progress visible to all stakeholders via SSE streaming