Application of Intelligent Grading and Risk Assessment Models in Commercial Platforms

The article introduces a quality‑grading model applied to commercial platforms, aiming to improve autonomous testing conversion and recall capabilities through objective data‑plus‑model evaluation.

Six core capabilities are required for large‑scale deployment:

Process control – unified workflow, visualized evidence, and integration of model triggering and verification.

Feature mining – extraction of both generic and business‑specific features, covering toolchain information, code analysis, coverage, etc.

Feature data collection – multiple ingestion methods (API, agent, remote, configuration, sync) to ensure fast and convenient data access.

Feature data storage & processing – unified data hosting, retrieval, tagging, and lineage management.

Strategy management – centralized strategy hosting, registration, training, debugging, scheduling, and result callbacks.

Annotation platform – visual feedback loop for model training samples, improving model accuracy and recall.

By tightly integrating these middle‑platform capabilities, the solution achieves configurable business onboarding, standardized test admission, unified data handling, standardized strategy development, and risk‑assessment feedback mechanisms.

The technical scheme includes:

Overall interaction flow: CI and testing stages collect feature data, persist it in a data platform, and trigger model evaluation via the process‑control platform.

Process‑control details: visualized feature extraction, real‑time data ingestion, model triggering, and post‑processing (pass, reject, retry) with annotation feedback.

Data processing: mapping source schemas, rule‑based transformations, multi‑level data retrieval (linkage, precise, broad, custom), and merging results into a wide table for model consumption.

Model aspects: classification problem (risk high/low), training using rule‑based, logistic regression, decision tree models, and online prediction via the data platform.

Configurable integration: separation of common and custom features, visual configuration, online review, and automatic model deployment without manual intervention.

Current results after two quarters show:

Hourly‑level module onboarding for over 20 business lines and 1,000+ services.

Model accuracy of 94% and recall of 90%.

Approximately 8% of test submissions converted to autonomous testing, significantly increasing delivery throughput.

White‑box analysis and incremental coverage contributed to a >1% recall of test submissions, surfacing 30+ bugs.

Future plans include exploring new risk‑assessment models, reviewing bad cases, deepening white‑box feature extraction, and supporting fully automated (unmanned) testing scenarios.