AI, for Less
I strongly advocate for "AI, for Less". My research shows you do not always need massive computational power — simpler, optimized models can yield exceptional results while remaining transparent and trustworthy. We emphasize reproducibility and simplicity over "black box" deep learning, and show that lightweight active learning, co-training, and warm-starts can match or beat expensive alternatives.
- FSE'26: Data-Light SE Challenge [pdf]
- VERIFAI'26: Exploiting Software Sparsity [pdf]
- TOSEM'26: LLMs Warming-Up Active Learning [pdf]
- CACM'25: The Case for Compact AI [pdf]
- TOSEM'24: Learning from Very Little Data
- EMSE'24: Co-Training for Defect Prediction
- IEEE Access'24: Partial Ordering for Model Reasoning
- EMSE'22: Minimizing Tech Debt Labeling Cost
LLMs & Deep Learning
Critical assessments of where LLMs and deep learning help — and where they hurt — in software engineering tasks. We study how to prompt, warm-start, and robustify LLM-based SE pipelines, and identify when simpler baselines still win.
- EMSE'26: Improving LLM Annotations [pdf]
- MSR'26: Domain Knowledge for LLM Opt. [pdf]
- IEEE Internet Comput.'23: Variances in Robust DL
- TSE'22: Oversampling for DL Defect Prediction
- MSR'22: Improving DL for SE Analytics
- MSR'22: GANs for Security Class Imbalance
Security & Vulnerability
Practical methods for vulnerability detection and cyberthreat intelligence, using active learning and adversarial ML to help teams prioritize security efforts without drowning in false positives.
- KAIS'25: Mining CTI Attack Patterns [pdf]
- ICDM'24: Temporal Cyberattack Patterns
- EMSE'22: Ensembles Against Evasion Attacks
- EMSE'21: Security Bug Report Classification
Defect Prediction & Software Analytics
Building better predictors for software quality, project health, and technical debt — using smarter sampling, transfer learning, and hyperparameter optimization rather than brute-force compute. Includes benchmark repositories for reproducible research across hundreds of SE tasks.
- MSR'26: Multi-Objective Optimization Repo [pdf]
- TSE'25: Is HPO Different for SE? [pdf]
- TSE'25: Static Code Mining Retrospective [pdf]
- SIGSOFT SEN'25: Replications & Negative Results
- ESA'23: Expert System for Cloud SE
- EMSE'22: Process vs. Product Metrics
- EMSE'22: Predicting OSS Project Health
- MSR'22: Stabilizing Models Across Projects
- ICSE'21: Early Life Cycle Defect Prediction
Accountability & Trust
AI software must meet the same correctness and auditability standards as any safety-critical system. We develop methods to verify, audit, and correct AI outputs — ensuring decisions can be inspected, reproduced, and challenged — and probe the fragility of causal reasoning in high-stakes environments.
- JSS'26: Explaining Optimization Heuristics [pdf]
- IEEE SW'25: AI in State Courts [pdf]
- EMSE'25: Causal Graphs in SE [pdf]
- EMSE'23: Interpreting Model-Based Opt.
- TSE'24: Equalized Odds via Pre-Processing
- IEEE Access'24: Scalability, Privacy, & Performance
