Team Jevo - Fine-tuned Jev with local computer use data

Browser agents need real interaction data in the same format they act in. So we collected structured data from daily browsing, then closed the loop to a local fine-tuned Jev-like model. Same schema at capture and serve time means the fine-tune actually transfers. Data stays on-device by default; the only optional cloud call is a short summary for intent labeling. The pipeline is repeatable: record more gold → re-export /v1/dataset → retrain Kev → compare against hosted Jev on the same harness.