Skip to content

Support Ticket Classifier

SageMaker models SageMaker model

This model package classifies a support ticket in Spanish, Portuguese, or English. It returns intent, urgency, sentiment, language, ranked candidates, and review flags. It runs on CPU in your SageMaker account. The weights are in the container. Inference does not download a model and does not call an external API.

It is not a hosted ticketing application. The inbox in the demo video is an illustration of one integration. Scores can be wrong. An unflagged prediction can be wrong. A person still has to review the ticket before you route it.

Delivery
SageMaker model package in your own AWS account

Quick start

Full guide →
  1. Subscribe
  2. Create the model
  3. Real-time endpoint
  4. Batch transform
  5. Clean up

What a response contains

category and intent are the same predicted label. The default labels are billing, technical_issue, account_access, cancellation, shipping_delivery, refund, complaint, sales_inquiry, and other. Urgency is low, medium, or high. Sentiment is negative, neutral, or positive. Language is es, pt, en, or unknown.

Each of intent, urgency, and sentiment includes a label, a score, a mode, and a ranked distribution. top_k holds the top three category candidates, or fewer when the label set is restricted. needs_review and review_reasons are a review heuristic. truncated is true when the tokenizer drops input past its 128-token window. model, version, and model_revision identify the build.

Scores are not calibrated probabilities. The model does not redact personal data.

Instance and price

The qualified CPU instance is ml.c5.xlarge, for a real-time endpoint and for batch transform. The public offer is USD 0.70 per instance-hour for both, and a 14-day software trial. Confirm the offer before you subscribe. AWS infrastructure charges are separate, including during the trial.

Enable SageMaker network isolation. The container does not need outbound network access. Serverless inference is not supported.

Walkthrough

Watch on YouTube. The listing says the interface is illustrative, the examples are fictional, and the model responses are from the published model.

Where to go next

Last updated November 11, 2026 · Suggest an improvement