StayCharted

PRODUCT TOUR · STAYCHARTED AMT

Your examples.
Your AI, at work.

See how AI Model Trainer turns labeled text and pictures into private, business-specific models.

Explore the walkthroughs ↓

See it in action.

Two step-by-step walkthroughs, with on-screen captions.

Screens show sample data; results depend on your own examples. Account details and some rows have been changed for privacy. Masking is a choice you make before training.

TEXT MODELS

From past decisions to new answers.

Follow a support-ticket model from upload to prediction, using 900 sample tickets across six categories.

01Start with the examples you already have

Upload a CSV, Excel or JSON file. Choose what to predict and review which columns the model will read. Keep useful information and leave out identifiers that do not help.

Choose the category to predict and the message column to learn from.View full-size screenshot ↗
02Review sensitive information before training

StayCharted AMT checks for supported types of personal and secret information. Choose to mask detected values, exclude affected rows, or keep them. Masking also applies to text processed later for predictions; detection may not find every sensitive value.

Detected email addresses and phone numbers with choices to mask, exclude or keep.View full-size screenshot ↗The review screen after masking email addresses and phone numbers.View full-size screenshot ↗

Masking does not change the original upload. Original text and picture ZIP training uploads are deleted after 90 days, unless deleted sooner. Some recorded screens show earlier retention wording.

03Choose how your text model learns

A Classifier Model learns word patterns. An AI Classifier Model also uses meaning from pre-trained AI. A Dedicated AI Model fine-tunes AI on your examples. Available options depend on your plan; review validation results to choose the approach that fits your data.

Model training options and the confirmation step before training.View full-size screenshot ↗
04Review results before publishing

See performance on examples held out from training, compare it with a baseline, and check accuracy by category. The sample report shows 92.1% accuracy on 127 held-out examples. It describes this demonstration, not expected results for your data.

Sample validation report: 92.1% accuracy on 127 held-out examples.View full-size screenshot ↗Accuracy by category, with the weakest categories shown first.View full-size screenshot ↗
05Correct labels and train a new version

Review the mistakes in the test set. If an example has the wrong label, correct it and retrain. Your published version continues answering until you publish its replacement.

Review predicted and expected categories and correct mislabeled examples.View full-size screenshot ↗
06Use predictions where the work happens

Try an individual record, fill a file with predicted categories and confidence scores, or connect your application through the Prediction API on a plan that includes API access. Review uncertain predictions before acting on them.

Try individual records and give feedback on recent predictions.View full-size screenshot ↗Support tickets with predicted categories and confidence scores.View full-size screenshot ↗Prediction API instructions with a placeholder API key.View full-size screenshot ↗

VISION · IMAGE MODELS

Teach it how you categorize pictures.

Start with a ZIP organized by category, or a spreadsheet of image links. Check near-duplicates, review your categories, then train a model to categorize new pictures.

Every picture is checked. AMT’s own copies are made from the pixels alone and carry no location data; your uploaded files stay as you sent them for the applicable retention window.

Explore image classification →

BUILT FOR TEAMS

Give the work a shared home.

Organize models, files and API keys by workspace. Invite team members as Builders, Reviewers or Viewers, and follow changes in the Activity log. There is no per-seat charge.

Read the walkthroughs

The overview

StayCharted AI Model Trainer in 90 seconds. Turn a column of text, or a folder of pictures, into a column of answers Upload examples your team has already labelled. Personal information is masked before anything is trained.

Pick a model: no AI, a pre-trained AI, or a Dedicated AI Model trained on GPUs. Every version is tested on examples it never trained on. See every mistake, and fix mislabelled examples.

Try it on a single record. Fill a whole spreadsheet, with a confidence on every row. Or call it from your own code.

With the Vision add-on, train on pictures too. New pictures, sorted, each with a confidence.

Invite your team as Builder, Reviewer or Viewer; team members are free. Every action is recorded in the Activity log. Build your own AI model. amt.staycharted.com

Train a text model

Build a support-ticket model, step by step. 900 labelled support tickets · 6 categories

Step 1 · Set up

Name the model and say what it should predict.

Step 2 · Upload

Upload the spreadsheet you already keep: CSV, Excel or JSON. It finds the column to predict, and counts the rows and categories. It measures which columns carry signal: the message is read, the ticket ID left out. Check the rows exactly as the model will read them.

Step 3 · Data quality

The data quality report: 853 usable rows across 6 categories. Personal information is found before training: email addresses and phone numbers. Mask it with one click. It is masked the same way in everything the model reads later.

Step 4 · Train

Choose the model. An AI Classifier Model reads meaning and trains in about a minute. Training runs in the background. You can close the page.

Step 5 · Validate

Tested on 127 tickets it never trained on: 92.1%, against an 18.1% naive baseline. Accuracy for every category, weakest first. Every mistake is listed. If the label was wrong, correct it for the next training. Publish the version to start using it.

Step 6 · Use it

Try a record and see the answer with its confidence. Fill a spreadsheet: upload 200 new tickets. All 200 rows filled, ready to download. Every row gets a predicted category and a confidence score. Or call the model from your own code with a workspace API key. If you need more, train a Dedicated AI Model on the same data, on GPUs, and compare. Your categories, learned from your examples. amt.staycharted.com

Sort pictures with Vision

Sort product photos with Vision. 120 product photos · 4 categories

Step 1 · Set up

Name the model and choose Pictures.

Step 2 · Upload

Add a ZIP with one folder per category, or a spreadsheet of picture links. The ZIP: caps, mugs, t-shirts and tote bags, 30 pictures each. The categories come from the folder names.

Step 3 · Check

Every picture is checked, and the copy AMT learns from is made from the pixels alone. Your uploaded file is kept as sent. Near-duplicate pictures are found, so one photo isn't counted twice. Keep one of each with a click: 110 pictures, ready to train. Look through every picture by category. Change a category or leave one out.

Step 4 · Train

Choose the AI Image Classifier Model. It trains in about a minute.

Step 5 · Validate

16 of 16 test pictures right. Results depend on your own pictures.

Step 6 · Use it

Sort the new arrivals: upload a ZIP of 24 new pictures. All 24 filled, ready to download. Every picture gets a category and a confidence. Sort pictures the way your team already does. The Vision add-on · amt.staycharted.com

YOUR NEXT STEP

Build your own AI model.

Start with the examples you already have. The Free plan includes access to an AI Classifier Model, and you can also use your own AI platform without training and compare its results.

Start free ↗