Inspiration: Wacky Wheels Co. keeps no parts inventory, so one late shipment, plant outage or bad part lot ripples through production, dealers, loaner cars and owners. We wanted AI agents that fix these problems without causing new ones, so we focused on acting only when we can prove the action is right.
What it does: A coordinator and three specialist agents read 15 operational events and handle each through the provided enterprise tools, following the policy handbook. They reschedule or transfer orders, notify dealers, assign loaners, hold lots, stop sales and contact owners. They do nothing on noise and escalate unconfirmed rumors to a human. Every event gets a graph path, a reasoning trace and a summary, shown in a Streamlit control tower.
How we built it: We used a Ground, Compile, Prove design: Read: a small local LLM (qwen2.5:1.5b through Ollama) and a rule reader interpret each message. Code turns names into ids and phrases into dates. Ground: a knowledge graph (about 17,000 nodes) decides which reading matches real records. Compile: the handbook is applied as exact code, with shared slot and loaner ledgers. Prove: the plan is dry-run on a copy of the systems and checked against 13 rules before any live call.
Results: On the 15 events: 879 tool calls, 0 failed, only the two rumors escalated, and no writes on noise. Replay, 31 tests and an independent audit all pass.
Challenges: Ambiguous dates like “next Friday”, name traps in the data, a small model that sometimes misreads a sentence, and compound events competing for the same slots and loaners.
What we learned: Let the LLM read and let code compute. Proving a plan before acting catches mistakes cheaply, and a small model is enough when a knowledge graph checks what it says.
