Modern-Go: A Low-Cost Multimodal Cashierless Retail Checkout System Using Edge Vision and Browser-Hosted VLM Attribution

Cashierless retail systems are usually presented as a hardware-heavy service tier, but the operational bottleneck is not only sensor coverage. The harder problem is turning noisy multi-camera observations into trustworthy checkout events without expensive ceiling grids or proprietary back-end services. This paper presents Modern-Go, a practical Amazon Go–style checkout stack built around compact YOLO detection, BoT-SORT tracking, CPU face recognition, historical frame buffering, and browser-hosted Qwen inference for ambiguous pickup attribution. The implementation is split across a Python vision subsystem and a TypeScript retail API, joined by local web bridges that keep video streaming, event attribution, and inventory/state management loosely coupled. We describe the architecture, the event pipeline, prompt design, and the store-product/customer data model used to support retail operations. On a staged test setup, the system sustains real-time camera streaming while limiting heavyweight vision-language calls to rare pickup/return events, which keeps the average per-frame cost low. The resulting design is not a research novelty benchmark, but it is a workable engineering blueprint for low-cost autonomous checkout with edge-friendly deployment characteristics.