FORGE is a hands-free voice copilot for industrial field-service technicians, built on Qwen's realtime omni model for the Global AI Hackathon Series with Qwen Cloud. One realtime session carries the technician's speech, spoken replies, live camera frames and all 25 tool calls, with replies reported under a second. The tools brief on a machine, look up parts and torque specs, navigate schematics and a 3D model, mark issues directly on the camera view, record measurements and guide procedures step by step. Every request is checked against the machine catalogue, so an unknown part is refused rather than guessed, and safety is enforced on the server instead of being delegated to the model: the lockout/tagout checklist requires a spoken confirmation for each individual item. A second model, qwen-plus, diagnoses faults in the background from fault data and measurements without interrupting the conversation, and alerts are raised when a measurement breaks a failure rule. At the end of a job the event log is turned into a work-order report and a shift handoff. The backend runs on Elastic Compute Service over HTTPS with OSS storage and a health endpoint that reports live status and the model in use.
Our takeThe pattern is worth naming: the model is allowed to be fluent, and the constraints live in the tool layer and the server. Refusing an unknown part number, or requiring a spoken confirmation per lockout item, is enforcement that survives a confident hallucination. Coverage is the open question - 25 tools and one catalogue are a small world, and the awkward case is a request that falls between two tools.