Intelligent Ti Machine Learning Platform: Industrial and Financial Applications

On September 7, the Cloud+ Community (Tencent Cloud official developer community) hosted a technical salon titled “AI Technology Principles and Practice” in Shanghai. Five Tencent Cloud experts presented the Intelligent Ti Machine Learning Platform (TI‑ONE) and its applications in industrial panels and finance.

The platform is a one‑stop solution for developers, covering the entire ML workflow: data preprocessing, feature engineering, model training, inference, and one‑click deployment. It supports traditional ML, time‑series, NLP, graph, and computer‑vision algorithms, and offers both private‑cloud and public‑cloud versions.

Key platform capabilities include:

Drag‑and‑drop task flow with built‑in components (PYtorch, Caffe2, Spark, etc.)

Automatic hyper‑parameter tuning (semi‑automatic, full‑automatic, Bayesian optimization, SMAC, genetic algorithms)

Data preview and visualization tools

Rich built‑in algorithms (distributed ML, deep learning, graph algorithms based on Angel, TensorFlow, PyTorch)

Notebook integration for custom code and model export

Full‑automatic ML pipelines with model monitoring, versioning, gray‑release, and A/B testing

Industrial panel use case : The platform was deployed for leading panel manufacturers (e.g., Huaxing, Tianma). Challenges such as high‑dimensional data, lack of data science expertise, and complex defect detection were addressed by:

Building a five‑layer solution (business platform, big‑data platform, AI platform, application scenarios, front‑end visualization)

Implementing anomaly detection, defect classification with multi‑stage modeling (object segmentation → clustering → classification), using models like PSPNet

Achieving real‑time monitoring, root‑cause analysis, and automated corrective actions, which reduced inspection time from 6 hours to 1 hour and improved yield.

Financial industry use case : The platform supported precise marketing, operation optimization, and risk control. A case study on corporate wealth‑product purchase prediction involved:

Data preprocessing (sampling, handling outliers, feature transformation, normalization)

Feature importance analysis (e.g., guarantee amount, owner age, employee count)

Model training with SMOTE for class imbalance and ensemble methods (Decision Tree, Random Forest, GBDT, XGBoost)

Achieving an F‑score of 86.3 % and AUC of 0.938, ready for production deployment.

The overall lessons emphasize the need for domain experts, data experts, analysis experts, and a stable modeling platform to turn data into business value.