QualAI is a two-year Italian national research project aimed at building a comprehensive framework to continuously monitor, assess, and improve the quality of ML-based software systems — from data and models to deployment and operations.

Why it matters
In 2020, a Google Health AI classifier for diabetic retinopathy achieved over 90% accuracy in the lab — yet failed in practice, discarding more than one-fifth of real hospital images and causing diagnostic delays of months. This case illustrates a fundamental truth: the quality of an ML-based system is far more than the accuracy of its model. QualAI addresses the full spectrum of quality challenges that arise throughout the lifecycle of AI systems, from training data and model design to integration, deployment, and live operation.

Research objectives

OB1 – Monitoring Framework: Define a shared knowledge base drawing on source code, notebooks, and SE-related data sources, providing the data foundation for all quality assessment approaches.

OB2 – Data & ML Model Quality:  Detect and mitigate quality issues in training data and ML models, covering robustness, fairness, privacy, interpretability, efficiency, and reproducibility — including computational notebooks.

OB3 – ML Integration Quality: Identify quality smells and communication gaps between data scientists and software engineers, address technology mismatches, and safeguard system security at the integration level.

OB4  – Deployment & Operations: Detect configuration smells in CI/CD pipelines and container images, monitor live systems for data drift, and automatically suggest corrective actions to keep deployed models reliable.

The research will be conducted in line with the principles:

⚖️ Cost-Effective Recommendations
Every recommendation ranks issues by the ratio of remediation cost to quality benefit, helping teams prioritise what matters most.
🔍 Explainable Outputs
Each recommender provides human-readable rationale — in text or visual form — so practitioners can understand and trust the suggestions they receive.
🔄 CI/CD Integration
QualAI slots into existing pipelines, triggering quality checks automatically on every commit throughout the development lifecycle.
🔬Empirical Validation
All proposed approaches are validated through mixed-method studies combining repository mining, surveys, and interviews with industrial practitioners.

Expected Outcome

QualAI will deliver a suite of approaches to assess and monitor the quality of an ML-based system across multiple dimensions — from data integrity to operational stability. All tools and datasets will be released under open-source licences.