MakerCheck

Finance teams run on spreadsheets that change every week, and the review is usually a person reading cells. MakerCheck treats a workbook change like a pull request. It compares the candidate .xlsx with the approved version, parses the formulas into a dependency graph, applies deterministic risk rules, and reports each newly introduced risk with its exact cells and the outputs downstream of it. A human checker then approves or requests changes, and the decision is logged with both workbooks’ hashes.

In the demo, an analyst adds a JPY position to an FX exposure workbook but doesn’t extend the total, and types in an FX rate instead of taking it from the approved Rates sheet. MakerCheck reports exactly those two risks and the summary cell they reach, and the corrected workbook comes back clean. The AI agent, Tencent’s CodeBuddy, only explains the findings and drafts the reviewer’s message: it can’t invent a risk or approve a workbook.

I built it for the Productivity Agent track of the AI Tinkerers × Tencent Cloud Agent Development Challenge in August 2026.