An operations system for device buyback, run by agents with approval gates.

In a buyback business, each traded-in device has to be checked, graded against what the customer said in the questionnaire, priced against the internal sheet and the live market, and then sent to the right channel. The system was built as a single place to run that pipeline, framed as an operating system in which AI agents propose and carry out the work while the operations team keeps the final say through approval gates.

The command screen is a streamed Claude chat grounded in the pipeline data. It answers questions, draws charts into a side canvas, shows device and lot cards, and runs real actions only after someone presses approve. Beside it sit the approval queue, the agent roster and an activity feed. The pipeline view is a master table with a stage conveyor and a drawer showing each device's journey, economics and condition report.

Grading uses a capture wizard that records vision findings, diagnosis checks and battery health for each device. Devices graded A to B become draft listings for the refurbished storefront; C to F go straight to the auction desk. A counter diagnostic screen uses the device camera with real vision grading, and the storefront and vendor auction portal receive published listings and lots live, so the whole route from counter to channel can be followed in one place.

The pricing desk sets the internal price sheet against live market prices and a manual floor for every device, and suggests whether each one goes through the app auction or a bulk email to vendors. Market comparisons are fetched server-side from a public trade-in site, and lot cohorts carry festive-season forecasts. The channels screen consolidates five spreadsheets into one auction file from a real upload, and the autopilot runs a vendor round, gathers replies and drafts the negotiation, all through the approval queue.

This is a client demo running on deterministic seeded data, with the client's own trade-in references, part codes and lot numbering, so the team can judge it in their own terms. It also includes a mismatch analysis between questionnaire answers and graded condition, built from the client's own variance data. There is no evidence yet of use on live operations, and no outcome has been measured.