Intelligent Automation Architecture
Self-healing test automation, LLM-based test generation, and an AI governance rail that lets every model call be audited. Papers D & E.
Predictive Analytics for Software Quality
Process-metric-based defect prediction at scale; cross-repository transfer; honest feature-importance analysis under the corrected dataset. Papers A & B.
Research-to-Production Translation
Risk-based prioritisation deployed as a real microservice on real CI/CD; vision paper on post-execution defect attribution. Papers C & F.
Research questions driving the program
The questions are deliberately narrow and falsifiable; nothing on this page is speculative.
How predictive are process metrics at scale?
Answered in Paper A: six classifiers, 296,457 instances, RF AUC 0.8998. The corrected file_age_days changes signal distribution; commit_frequency jumps from near-zero contributor to a meaningful secondary signal.
Do those models transfer across projects?
Answered in Paper B: yes, with caveats. Cross AUC 0.867, cross F1 0.631, and an AUC–F1 asymmetry that practitioners must account for when using the model for ranking versus binary classification.
Does prediction translate to operational lift?
Answered in Paper C: top 10% of files capture 43.82% of defects — a 4.37× risk-vs-random lift and 81.4% of the oracle ceiling at the 18.61% base defect rate.
Can locator failures be auto-recovered?
Addressed in Paper D: tree-ensemble ranking over eight DOM feature families across 2,400 mutation events. Confidence gating decides auto-heal vs. flag-for-review.
Can LLMs replace manual test authoring?
Addressed in Paper E: 312 reqs / 3 domains, 68% effort reduction, 94.1% req coverage, 96.3% first-pass verification — with deterministic rule verification as a non-LLM safety net.
Can we attribute failures, not just predict them?
Paper F outlines the fusion of dynamic failure signals with SHAP-explained priors, producing triage suggestions instead of ranked file lists.