Decoding sequences across language, biology & security.

Bengali emotion detection lacked a clean, manually-verified corpus spanning the language's core emotional categories.

Built the Robust Bengali Emotion Corpus (RBEC) and benchmarked multiple model families against it for six-way emotion classification.

Fine-tuned XLM-R reached state-of-the-art accuracy with strong agreement scores, setting a new benchmark for the task.

Farmers describe crop ailments in colloquial Bengali text, a format poorly served by existing agricultural diagnosis tools.

Curated a custom dataset of colloquial disease narratives and evaluated a multi-architecture classification system across it.

Demonstrated that disease identification from informal, real-world farmer language is tractable at scale.

Religious hate speech in Bengali is widespread online but under-resourced for automated, reliable detection.

Developed a novel corpus of real-life statements and fine-tuned Bangla-BERT for binary hate speech classification.

Achieved near-ceiling classification performance, offering a strong foundation for moderation tooling.

No standardized benchmark existed for binary political statement classification in Bengali text.

Curated a dedicated dataset and evaluated 20 models, from traditional ML to Transformer-based systems, head-to-head.

Established a reusable benchmark and clarified the performance gap between classical and Transformer approaches for this task.

Cloud network anomaly detection must contend with severely imbalanced traffic data and opaque model decisions.

Proposed CAPGBN, a Cross-Attention-Based Probabilistic Graph and Bayesian Network framework for risk assessment.

Validated the hybrid approach against a highly imbalanced dataset of real network session records.