AIRIMF: An AI-Driven Risk Identification and Mitigation Framework for Software Project Management

Project managers have long relied on periodic risk reviews and intuition-based judgement to handle software issues—a strategy that simply does not scale with the pace of modern Agile development. This paper introduces AIRIMF, a machine-learning pipeline designed to flag high-risk tickets automatically at the moment they are created, before any development work begins. We collected 2,593 bug reports from microsoft/vscode and 744 from facebook/react, applied SMOTE to balance the training data, and trained a Random Forest classifier on five features that are strictly available at submission time: title length, description length, initial comment count, and two sentiment scores derived from the title text. After carefully removing every feature that could introduce target leakage, the model achieved 76.21% cross-validation accuracy on VS Code and 83.37% on React, with Class 2 (Challenging) recall of 82% and 89% respectively. A McNemar test confirmed that the improvement over SVM is statistically significant (p < 0.0001). SHAP analysis showed that description length is the single strongest early risk signal, while sentiment features contribute a meaningful six percentage-point gain when combined with structural metadata. We further evaluate cross-project calibration on the Angular repository and discuss how Platt scaling can recover accuracy when a model transfers to a new codebase with a very different class distribution.