Identify parts from a photo: your team's AI model on a technician's phone

The work is split between the office and the field.

In the office, once:

1. Collect photos of each part your team already knows. A folder per part number works, or a spreadsheet of photo links with a part number column. Phone photos taken the way technicians will take them are best.
2. Train a model in StayCharted AI Model Trainer (AMT). An AI Image Classifier Model uses a ready-made vision AI and learns your part numbers from your photos. AMT tests it on photos it never saw and shows accuracy per part, and how often its sure answers are right.
3. Publish it, and share it. Team members sign in to the app with their AMT account. For technicians who don't have one, a Builder makes an access code for that one model and hands out the code or its QR code.

In the field, every time:

1. Open the app, pick the model (or scan the code once), and tap Take a photo.
2. Read the answer: the part number, large, with the model's confidence, such as "92% confidence".
3. If the model isn't sure, the app says "Sent to your team to check". The photo lands in the Review Queue, where someone who knows the part confirms or corrects it. That answer becomes training data for the next version.

The Review Queue is available on Essentials and Business. Picture models require the Vision add-on.

We measured it in our part identification benchmark, with a hard case on purpose: 100 LEGO part numbers, many of them nearly identical bricks and plates, from a public photo set (Gdańsk University of Technology, CC BY). Each model trained on 20 photos per part and was tested on 1,000 photos it had never seen.

The more useful number for a technician is how often the sure answers are right. At its default cutoff of 0.5, the AI Image Classifier Model was sure of 32% of the photos, and 322 of those 323 answers were right (99.7%). At a cutoff of 0.3, it answered 65% of photos and was right on 98.2% of them.

What that means on site: set the cutoff where your team is comfortable, take the sure answers, and send the rest to a person. Reviewed photos can train the next version; measure again to see whether accuracy and coverage improve. Your own parts will score differently from LEGO bricks, so AMT measures your model on your own photos before anyone relies on it.

Many technicians, contractors and branch staff will never have an AMT account, and shouldn't need one to identify a part. An access code gives a phone one model and nothing else: no training data, no other models, no files, no settings.

A Builder sets the limits when making a code:

• Expiry: 90 days by default, a set number of days, or never.
• Phones: 10 by default; each phone that enters the code counts.
• Answers a day: 500 by default, across all its phones.

Codes can be paused, turned back on, closed to new phones, or revoked so every phone stops at once. A single phone can be removed. The workspace Owner can switch codes off altogether. Every code that's made, paused or revoked, and every phone that joins or is turned away, is in the workspace's Activity log, and each answer records which code asked. Access codes come with the Prediction API, on Essentials and Business.

• Photos are resized on the phone and sent only to answer the question.
• They aren't kept, except when an unsure answer goes to the Review Queue for your team to check.
• The app doesn't ask for or send the phone's location, and it has no ads or tracking.

The same app works for any model your team publishes. Field teams use photo models to:

• grade condition of an asset the way your inspectors already do, photo after photo, for condition-based monitoring;
• spot damage or defects, as in our vehicle damage and visual inspection work;
• look up a category by typing: a returns reason, an issue type or a product category, from a text model.

We'd rather say it plainly:

• No answers without signal. The app needs a connection to send a photo.
• It doesn't read the label. If the part number is printed on it, the model looks at the photo, not the text, for now.
• One answer per photo. It doesn't list several likely parts yet.
• No training in the app. Models are trained and published in AMT on the web.
• Android only, for now. The iPhone app follows Apple's approval.

1. Start free and upload photos of the parts your team already knows, or connect the storage folder they're already in.
2. Train an AI Image Classifier Model and read its test report: accuracy per part and the confidence table.
3. Publish it, set the review cutoff, and make an access code for the technicians who'll use it.
4. Install the app from Google Play and take the first photo.

Picture models need the Vision add-on on Essentials or Business. See pricing and the mobile app page.