Criteria move from people's heads into scenarios
Each line writes its criteria down as text and OK and NG reference images. When people change, the criteria do not.
Turn inspection by eye into decisions you can look up
Set a part on the tray and the tablet shoots the moment it holds still; the vision model on your server answers OK or NG with the defect type and a finding, using the criteria and reference images in the scenario. It runs in a browser, nothing to install.
What changes
Every decision is kept with its image
What was called NG and why — image, finding and confidence — is filed by date. You can find the images from the day a problem started.
It fits the line better the longer you use it
Results people confirm become training data for a model dedicated to that line, and that model takes over judgment as far as its record earns.
It shoots when the part stops and answers at once
The operator only sets the part down. Motion detection and blur rejection run on the tablet, and the result is shown large enough to read from across the station. A manual shutter is there too.
- An NG is sorted into the defect types the scenario declares — scratch, dent, discoloration — with a finding that says where.
- When the vision model cannot be reached it does not fall back to OK; it shows "unable to decide — check by hand".
It stops the wrong part
The operator enters the expected part number and the capture is answered as a match or a mismatch. When the tablet can read the label's barcode the code is compared first; without one, the vision model checks the part against the number.
Nothing is lost when the network drops
Captures are saved on the tablet first. Those that cannot reach the server are sent in the order taken once it is back, each judged only once, and the record keeps the time they were taken.
Supervisors check it in the console
Find results by date and scenario. When the vision model and the dedicated model disagree, or confidence is low, a person settles OK or NG from the review list.
- See the day's summary and counts by defect type, and download results as CSV.
Judgment that gets better with use
On day one the vision model judges. Once enough human-confirmed images of each type are in, the console says it is time to build a model dedicated to the line. The training data, selected by rule, is downloaded through the API.
- The dedicated model judges alongside the vision model, and disagreements go to the review list.
- It starts by being recorded only; as its record earns it, a supervisor hands it judgment one stage at a time. When it starts to miss, the system steps it down on its own.
- RoadmapWe are building a memory of confirmed cases, so similar parts are judged at once without waiting for training.Need this? Tell us
Roadmap
One engine, many inspections
Beyond still images, we are building the same judge-and-record loop for fixed cameras watching the line. Early fire and smoke warning is on the same path.
Video is processed by an edge device beside the line, which sends the server only the events to judge and the crops that go with them, holding them while the connection is down.
Counting
Counts items as they cross a line on the conveyor, beside the work order's plan. Items that pass overlapped are checked again.
Need this? Tell us
Hard-hat watch
Watches for hard hats on workers inside a zone. It alerts only when a miss lasts several seconds, not on a single frame, and it never judges faces.
Need this? Tell us
Anomaly detection
Starts without any defect images. It marks where a part differs from normal ones, and the vision model explains what is different.
Need this? Tell usIt connects to the systems on site
Each decision is posted to systems such as an MES as a signed event (CloudEvents) over a webhook. The API is described by a published OpenAPI document, so integration code can be generated from it.
- RoadmapWe are building intake of work orders, lots and expected part numbers from the MES, for the wrong-part check to use directly.Need this? Tell us
- RoadmapWe are building MQTT and OPC UA output for results.Need this? Tell us
It runs on your server
Images stay in the plant
It judges with a model served from your own GPU server through an OpenAI-compatible API. Inspection images are not sent to outside services.
No database server to run
Decisions and images are kept in folders on the server. A backup is a folder copy.
Inside what you already run
Run it as a standalone server, or mount it in an ASP.NET Core app as a NuGet package.
How to start
Pick one inspection
Choose one inspection where it would matter most, and write its criteria down as text and a few OK and NG reference images.
Run it with one tablet
Put a tablet over the tray and judge on the real line. At first, people check the results.
Measure it from the record
See in the console how often the decisions matched what people confirmed.
Hand over and widen
As data builds up, train a dedicated model, hand judgment over, and extend to other lines and inspections.
Tell us what you need to inspect
Even if it is not on this page. Tell us what parts, against what criteria and how many a day, and we will find a starting point together. The roadmap's order is set by requests like yours.
For developers
Open source under the MIT license. Setup instructions, the API document and the package are public.
What it does not do
- It does not promise accuracy up front. Judgment quality depends on the part, the model and the criteria, so it keeps human confirmations and decision metrics for you to measure it on site.
- There are no user accounts. It assumes one organization runs it: the operator screen is open, and only configuration changes need the admin PIN.
- This site sets no cookies. There are no tracking scripts either.