Developed an LLM-powered program management agent at Hendrickson that integrates with existing project tooling to automate task tracking, proactively surface blockers, and generate status summaries. The agent reduces manual coordination overhead for program managers and provides real-time project health insights through a conversational interface backed by FastAPI and PostgreSQL.
- Role
- AI engineering intern
- Challenge
- Program status lived across tools, making blockers easy to miss and summaries expensive to assemble.
- Outcome
- Built a working agent that surfaces blockers, tracks tasks, and generates project health summaries.
Decisions
- Ground responses in existing project data
- Separate retrieval from generated summaries
- Keep the interface conversational while preserving source context
Architecture
- Project data
- Retrieval layer
- Agent workflow
- FastAPI service
- Program manager interface
What changed my thinking
- Useful agents need narrow authority and visible provenance
- Fallback behavior is part of the product
PythonLangChainFastAPIClaude APIPostgreSQL