AI Platform‑Powered Automated Ticket Routing: Modeling Workflow, Feature Engineering, and Intent Recognition

With the rapid development of artificial intelligence, many enterprises are using AI to improve efficiency, reduce costs, and enhance competitiveness. This article describes how the Haro AI platform can be leveraged to automate the routing of customer‑service tickets, using a case study of the "Universal Ticket System".

Modeling Process

The AI platform follows a four‑step modeling pipeline: feature processing, model training, model evaluation, and model deployment. Each step includes sub‑tasks such as feature storage, selection, cleaning, and transformation; environment setup, hyper‑parameter tuning, and training acceleration; performance, time, and hyper‑parameter evaluation; and deployment of TensorFlow, PyTorch, PMML, or GPU models.

AI Platform Solutions

Feature Processing – The platform supports online feature storage (Hive, Kafka, RocketMQ → HBase, Redis, RocksDB) and feature transformation operators such as Normalization.

Model Training – Users can configure AutoML or expert mode, select models (e.g., IntentRecognitionNNIAndRay built with ALBERT), and apply preprocessing, hyper‑parameter optimization, and distributed training.

Model Evaluation – Visualization of hyper‑parameter search (via NNI), automatic and manual evaluation, and performance diagnostics are provided.

Model Deployment – Supports TensorFlow, PyTorch, PMML, GPU models and Faas‑style deployment to application clusters.

Intent Recognition Algorithms

The platform uses BERT and its lightweight variant ALBERT. BERT is trained with Masked Language Modeling and Next Sentence Prediction to capture contextual semantics. ALBERT reduces parameters and memory usage while maintaining performance, making it suitable for scenarios with limited training data.

Universal Ticket System Use Case

The original manual workflow required customer‑service agents to route tickets and developers to resolve issues, leading to low accuracy and inefficiency. The new automated workflow replaces manual routing and problem‑solving with AI‑driven classification and automatic handling.

Automatic Routing Logic

Customer descriptions are labeled with tags and confidence scores; tickets with confidence above a threshold (e.g., 0.7) are automatically assigned to the appropriate development team.

Automatic Processing Logic

For certain tags, the system invokes predefined APIs to resolve issues without human intervention.

Integration Process

1. Requirement submission by backend developers. 2. Feasibility assessment by the AI platform. 3. Model development or reuse. 4. Model evaluation and operator (算子) consolidation. 5. Documentation of the operator. 6. Backend developers integrate online inference according to the documentation.

Summary

The AI platform enables configuration‑driven feature storage, operator‑based feature processing, and fully automated model training, evaluation, and replacement. Users without deep algorithm expertise can adopt AI solutions, while experienced engineers can fine‑tune models via hyper‑parameter optimization and visualization.