All work
Alenmaa Hackathon · SDAIA Academy · 2026
Lessons learned for PMOs
An AI assistant that helps project teams find and reuse lessons from past projects, and asks a person before saving new ones.
Built for the Alenmaa Hackathon and the SDAIA Academy program, 2026. I designed the agentic RAG system and built its knowledge base.
- LangGraph
- Agentic RAG
- Streamlit
- Retrieve
- Route
- Reflect
- Approve
What I built
- A LangGraph workflow (retrieve, route, then search or consult, then reflect).
- A Chroma vector store with local embeddings (bge-small-en-v1.5), using a Claude model through OpenRouter.
- A Streamlit app with Search and Consult modes.
Key decision
New lessons go through a human approval gate, and the reflection step defaults to not saving. This keeps weak lessons out of the knowledge base.
Data
NASA’s public lessons-learned records plus synthetic documents in a Saudi project context.
What I learned
The first reflection prompt was too permissive and let weak lessons through. Tightening it, and making the step default to not saving, fixed that.