Research

What if machines could explain themselves?

When one of our models fails, it can tell you where and why. Our networks sense their environment and adapt on the fly. Our drones find their own way. We build systems with a kind of self-awareness — and every project here starts with a question nobody has answered yet. Maybe yours.

Completed project · 2026

EV-Drift · Detecting data drift in license plate recognition

A license plate recognition model can fail silently when image conditions change: a dirty camera, a shift in lighting, bad weather. No error is raised; the plates just stop being read. EV-Drift is a proof-of-concept system we built to catch this drift in batches of electric vehicle images, judging each batch by the question that matters most in operation: can the plates still be read?

Every image passes through a recognition-verified plate reader (blue-plate localization plus a CTC-based CRNN) and is sorted into five readability categories. At batch level, five signals (unreadable share, plate blueness, reading confidence, internal features and image statistics) are normalized against normal batches of the same size and fused into a single drift score. When an alarm fires, the system matches the pattern of signals against known degradation types to estimate the likely cause and recommend an action.

Two research modules extend the detector: minimum-eigenvalue noise level estimation and a metadata-only detector that works without the recognizer. The complete system runs as an authenticated web platform with evidence panels, a recovery grid, an experiment browser and a simulator.

Drift detectionAUROC 0.995 on calibration degradations, 0.938 on degradations never seen in calibration
Metadata-only detectionAUROC 0.997 / 0.937, no recognizer required
Noise estimationCorrelation of 0.997 with injected noise
Benchmark24 degradation types at 5 severities; audited test set of 6,047 images
Platformdrift.ccnlab.net (authenticated access)

Ongoing project

CADI · Causal Attribution for Drift Intervention

When a deployed model starts failing, something in the world has changed — but usually many things changed at once, and only one of them matters. CADI is a monitoring system that identifies which change is actually hurting the model, and repairs only that one. Without labels, and without ever making the model worse.

Research areas

Trustworthy & causal machine learning

Drift diagnosis, causal attribution, targeted repair of deployed models, and guarantees on where a failure originates.

driftcausalityattribution

Next-generation wireless networks

MAC and routing protocols, resource allocation, and spectrum access for dense, heterogeneous, and cognitive radio networks.

MACroutingspectrum

UAV localization & aerial networks

Positioning and trajectory planning for aerial platforms used as mobile sensors and as flying relays for ground networks.

localizationUAV relays

Security for IoT & smart infrastructure

Federated and learning-driven intrusion detection, anomaly detection at the edge, lightweight cryptography for constrained devices, and quantum-resistant consensus for distributed sensing.

federated learninganomaly detectionlightweight crypto

Deep learning for health & vision

Multimodal medical imaging — from silicosis detection on radiographs to physiological signal estimation such as respiratory rate and cuff-less blood pressure, with calibrated uncertainty.

medical imagingbiosignalsuncertainty

Intelligent optimization & metaheuristics

Nature-inspired optimizers and their enhanced variants for global optimization, feature selection, and model tuning — the search engines behind many of our learning systems.

metaheuristicsglobal optimizationfeature selection

Resources & code

coming soon

Code releases accompanying our papers will appear here.

github.com/…

Datasets and benchmarks.

coming soon

Tools, demos, and tutorials.

coming soon