Grab Engineering presents a feedback-driven framework for verifying knowledge graph (KG) relationships at scale in dynamic domains like food delivery. The system treats unverified LLM-generated graph edges as hypotheses, injects them into live search results using an exploration vs. exploitation strategy (multi-armed bandit), and uses weighted user interaction signals (clicks, purchases, dwell time) to promote or prune candidate edges. The four-step loop covers hypothesis generation, candidate injection, signal aggregation, and graph topology updates. A food delivery case study illustrates how a candidate parent-child link like 'Noodle Soup → Dry Mee Pok' gets validated through real user traffic without manual curation. Future work includes hierarchical confidence tiers with a state machine (proposed → shadow → canary → production) and adversarial/spam resistance mechanisms.
