Skip to content

DATA FLOW / PILOT LIMITATIONS

Your data.
The pilot's limits.

Understand how the product website and desktop pilot work, and what the current verification covers.

01 / THIS WEBSITE

Explore without uploading photographs.

This product website does not upload your PCB photographs or run cloud analysis jobs. It shows product information, a captured desktop screenshot, and a synthetic demo. There is no account registration form, email collection field, or subscription checkout on the site.

The synthetic demo archive is a small example project. Downloading it does not require sharing a photograph of your board. The browser demo displays illustrative content; actual board reconstruction runs in the desktop application.

Opening the site creates ordinary web requests between your browser and the hosting service. Hosting traffic logs are a separate data flow; this page makes no unverified promises about logging or retention periods.

02 / DESKTOP AND SHARING

Local work and synchronization are separate actions.

Local project
Photographs, edits, and analysis results stay in the project folder on your computer. Local ONNX models run on your computer. Local analysis, editing, and KiCad export require no account or network connection.
Team sync
An upload you explicitly start in the desktop application sends the selected project's photographs and revision to the configured team service. Team members can access those resources according to their roles. Using local analysis does not prevent separately synchronized photographs from being stored on the server.
Cloud suggestions
This is a separate, optional action. Selected image crops are sent through the configured server to an OpenAI vision model. Returned component or connection interpretations are saved as suggestions; they never automatically become confirmed electrical connections.
Project archive
A shared .rapcb archive may contain photographs and project information. Before sending an archive to another person, check its contents and whether you have permission to share the board photographs.

Uploading should be an explicit choice. For confidential or company-owned boards, assess local work, team uploads, and third-party cloud suggestions separately. This website does not start those actions on your behalf.

The desktop pilot's team service, storage terms, and provider account depend on the deployment. This page does not guarantee one retention period or data region for every installation. Revoking a user's server access does not remotely erase project copies that user has already downloaded.

The application sends cloud requests with store=false. This setting does not replace the provider's own data processing and retention terms. Review the relevant provider account and privacy terms before enabling cloud processing.

03 / ACCURACY AND SCOPE

A photograph cannot explain the entire board.

  • Inner layers and hidden traces: ordinary front/back photographs do not reveal inner copper layers, the underside of a BGA, or connections covered by components.
  • Component identity: a detected box and category do not establish an exact manufacturer part number, component value, or pin function. Unmarked or unreadable values remain unresolved.
  • Netlist: proposed, unknown, and rejected connections are excluded from electrical nets. Visually observed copper continuity is not an independent electrical measurement.
  • Physical dimensions: a photograph without calibration against a known distance does not produce a physical PCB file; a schematic and netlist draft can still be exported.
  • KiCad checks: ERC/DRC checks assess file and design consistency. They do not independently prove equivalence to the photographed board or readiness for fabrication.

The current model measurement

The component model's class-averaged score is 0.1654 AP50 on its limited held-out test group. This measures bounding-box/category detection, not an overall accuracy percentage or electrical netlist success rate. The model misses and misclassifies components; small components especially require manual review.

There is no manually verified PCB-marking OCR evaluation or real-board electrical net-accuracy benchmark. The synthetic demo and file-consistency checks do not substitute for those missing measurements.

04 / CURRENT AVAILABILITY

The verified scope of a private pilot.

The Windows desktop workflow has been verified with real local model inference, project save/open, and KiCad draft export. Linux packages have not yet been built or run. A configured team server, live OpenAI requests, and a test subscription checkout are not offered as running services on this product website.

No public desktop installer is available on this site. The downloadable content is a synthetic demo archive. The site promises no certified recognition performance on real photographs or certified outcome for commercial use.

05 / IMAGE AND MODEL SOURCES

Image source and attribution.

The board photograph shown in the site's desktop screenshot comes from the PCB Component Detection Dataset and YOLO Model dataset. The photographs and annotations were published by the dataset contributors under Creative Commons Attribution 4.0. The source license is available in LICENSE-DATASET. The photograph is scaled within the desktop view and displayed with the application's analysis annotations.

These dataset labels do not verify exact part numbers, manufacturer pin identities, or circuit connectivity. The RA PCB pilot component model was trained on the licensed photographs using the original Apache-2.0 YOLOX-S architecture. Local OCR uses the official PaddleOCR PP-OCRv6 small models. The YOLOv8 checkpoint supplied in the source dataset is not used as RA PCB's model.

The downloadable SYNTHETIC_DEMO.rapcb is an explicitly drawn example project, separate from the real-board photograph dataset. Illustrative content and real-image candidates are identified separately on the site.