Anomaly Detection & Handling

Operational decision systems for rare events, defects, domain shifts, biomedical monitoring, autonomous systems, and manufacturing quality.

Focus

My anomaly-detection and handling work focuses on operational settings where models must support review, escalation, or deployment decisions under class imbalance, noisy measurements, scarce labels, domain shift, or limited historical data.

Core Themes

• Rare-event detection with explicit false-alert, purity, sensitivity, and missed-event tradeoffs.
• Automated inspection under scarce and imbalanced defect labels.
• Manufacturing shape-error modeling and early design-phase risk simulation.
• Biomedical monitoring from label-free or noisy flow-cytometry signals.
• Continual-learning decisions for segmentation systems that must adapt after deployment.
• Domain-shift and deployment-readiness framing for operational ML systems.

Representative Work

• Deep CNN-based automated optical inspection for aerospace components.
• Object shape-error modeling and simulation using morphing Gaussian random fields.
• Deep learning-enabled detection of rare circulating tumor cell clusters in whole blood.
• Label-free flow cytometry of rare circulating tumor cell clusters in whole blood.
• PILOT: data-free continual learning for real-time semantic segmentation via boundary guidance.

Decision-Science Connection

The key output is not only an accuracy number. It is a handling decision: classify, review, escalate, retrain, monitor, or defer because the evidence is not yet reliable enough.