-
Full audit trail, including all tool calls and actions performed.
-
Top of Events page, which shows each event and what plays out in depth.
-
Ontology page, showing the graph.
The problem: a just-in-time automaker with no parts buffer, where one disruption ripples into schedules, dealers, loaners and owners.
How it works, in five steps: understand the message, ground it in the knowledge graph, act through the tools via four agents, hold back on noise and rumors, then explain every decision.
A results table with the run's real numbers: 15/15 events on the independent checker, 879 calls, 0 failed, 50 orders moved, noise untouched, rumors escalated.
The design idea that sets the project apart: the LLM only extracts phrases, while code owns IDs, dates and policy. That gives reproducibility (an identical rerun, and the same actions with the LLM switched off) and a very small token bill (about 9,000 tokens for the whole run).
Real challenges from this build, including the robustness test that caught a bug in our fallback, which shows honest engineering to judges.
Built With
- claudecode
- geminiapi
- python
- vscode
