Parts classification with AI: from a photo, a description, or both

In both, the model learns your answers, not a generic list. And in both, the uncertain ones go to a person, whose answer trains the next version.

A technician has a worn part in hand and no readable label. They take a photo; the model suggests the part number. The clear cases come back at once. The uncertain ones go to the parts desk.

What you need:

• Photos of each part, about 20 or more per part number, including real phone photos taken where the work happens, not only clean catalog shots.
• Your most-requested parts first: start with 100 to 200 part numbers, or a smaller set that covers your frequent requests.
• Photos sorted by part number, in folders (parts/PN-10423/…), in the file name (PN-10423_front.jpg), or in each photo's keywords. They can stay in your Amazon S3 or other storage. See classify images in an S3 bucket.

What it's good at: parts with a distinctive shape. What's hard: parts that differ only by size, with nothing in the photo to show scale, and parts that look identical but differ in material or rating. For those, add text (below), or a scale in the photo.

Text is where most parts data lives:

• Supplier catalogs and new items: "BRG DEEP GRV 6205-2RS C3", to be placed in your part families or a standard like ETIM or ECLASS.
• Master-data cleanup: thousands of old records with no class, the wrong class, or the same part under three descriptions.
• Purchase lines and spend: what was bought, for spend classification or a UNSPSC code.
• Work orders: what failed and which part was replaced, for maintenance work orders.

What you need: an export of parts your team has already classified, one row per part: the description columns (short text, long text, manufacturer, supplier category) and the class your team gave it. A model can read several columns together.

Abbreviations, typos and supplier shorthand are normal. The model learns them from your own records, the way a new colleague would.

Some parts can't be told apart by eye: two bearings of the same size with different seals, two fuses with different ratings. Others can't be told apart by text: a description that says only "bracket". Given a spreadsheet of photo links, a picture model can also read a text column with each photo: a technician's note, the text on the label, the asset it came off. The photo narrows it down; the text decides.

All on StayCharted's production service, on public data, scored on items the models never saw.

From a photo: 100 look-alike parts. On a public set of photographed LEGO parts (100 part numbers, 20 phone photos each, many that differ by one stud):

The mistakes were the look-alikes a person would mix up: a 1×4 plate taken for a 2×8 brick. The full method: part numbers from a photo.

From a photo: car parts. Naming which of 21 exterior parts a cut-out car photo shows (bumper, mirror, wheel and so on): 89.9% right with the Dedicated AI Image Model, and most mistakes were front and back wheels, which look the same out of context. Car photo benchmark.

From a description: a store's catalog. Placing 2,844 new products into 158 of Wayfair's categories from the name and description: 95.3% right with the AI Classifier Model, 97.3% with a Dedicated AI Model. E-commerce categorization.

From a description: customs codes. On HS codes, exact 10-digit codes are much harder than the broad levels: in our HS code benchmark, the same model was right on 25% of exact codes but 74% at the 2-digit chapter. Classify at the level you need.

• Start with the AI Classifier Model (text) or AI Image Classifier Model (photos). They trained in under a minute in these tests and are a useful first step.
• The Dedicated models aren't always better. A Dedicated model trains the AI itself on your data, on a GPU. On the store catalog and on car damage it won. On 100 look-alike parts with 20 photos each, it did worse (75.3% against 87.8%). Our reading: with many parts and few photos each, a ready-made vision AI plus a classifier learns more from less.
• Test both on your own data. Every model is tested on items it never trained on, and the report shows accuracy for each part or class, so you choose on numbers.

1. Pick one question and one list: "which of our 150 most-requested spare parts is this photo", or "which of our 40 part families is this supplier item".
2. Export what you've already decided: the photos or the records, with the answer your team gave each one.
3. Train, and read the report. Look at which parts or classes it mixes up: those are the ones that need more examples, or a clearer rule.
4. Choose a review cutoff from the report. With many look-alike parts, a model can be right without being very sure, so a lower cutoff than the default may suit you.
5. Put it to work: fill a file or a folder of photos, call it from your systems through the API, or ask it from an AI assistant. Corrections in the Review Queue become training examples for the next version.

• Name a part it has never seen. A model knows the parts it has photos or records of. Matching a photo against one catalog image per part is a different kind of tool (visual search).
• See what a photo can't show: material, rating, thread pitch. Add text, or measure.
• Decide for you on the borderline ones. Those go to a person, on purpose.
• Show the top few candidates yet. It gives one answer and a confidence today.