Inspiration Company brains remember what was decided, but not when a decision was replaced. Teams keep building on outdated choices because nothing tells them the decision changed.
What it does Lineage tracks how decisions evolve. Ask why something is the way it is, and it walks the chain from the original decision through the incident, the debate, and the new decision, citing a source for every claim. Its contradiction radar checks a new PR, draft, or message against every active decision and flags work that relies on a superseded one, then suggests a fix. A decision timeline shows each decision's full history.
How I built it Every document is ingested into HydraDB with a hand verified decision graph through bring your own graph, using edges like SUPERSEDES, CAUSED BY, DECIDED BY, and MOTIVATED BY. Questions run through HydraDB graph context retrieval in thinking mode, and build string turns chunks and graph paths into LLM context. The app is Python and FastAPI with a lightweight web UI.
Challenges Making the graph deterministic for a reliable demo, which is why we declared it explicitly instead of relying on extraction.
What we learned On a small corpus, plain retrieval answers questions just as accurately. The graph adds what plain retrieval cannot: an explicit, citable chain of how a decision evolved, and proactive detection of work built on superseded decisions.
What's next A larger corpus benchmark with a tight retrieval budget, live Slack and GitHub connectors, and per person memory so engineers and managers get different emphasis.
Built With
- fastapi
- hydradb
- javascript
- python
Log in or sign up for Devpost to join the conversation.