Selected work

Representative Papers

Explainable AI for Screening Abuse-Related Trauma in Bangladeshi Children: A Training-Free Multimodal Framework Evaluated on Noise-Aware Synthetic Data

Jul 04, 2026

This study addresses the critical gap in Bangladesh’s child mental health landscape: the absence of localized, interpretable tools for early psychological trauma screening and severe shortages of specialized mental health resources. To tackle this, the authors propose ShishuRaksha—the first training-free, multimodal AI decision support framework designed for low-resource settings. It integrates four modalities—structured questionnaires, Bengali-language text, House-Tree-Person drawings, and facial emotion cues—employing clinically weighted fusion and cross-modal attention mechanisms to generate bilingual, interpretable reports linked directly to the national child protection system. Key innovations include single-modality coverage rules, perturbation-based additive attribution explanations, tree-ensemble surrogate modeling, and a noise-aware synthetic data benchmark. Evaluated on 500 synthetic cases, the model achieves an AUC of 0.874, significantly outperforming the Strengths and Difficulties Questionnaire (SDQ) unimodal baseline (AUC = 0.756), with ablation and calibration analyses confirming its robustness.

0 citationsRead paper

HaorFloodAlert: Deseasonalized ML Ensemble for 72-Hour Flood Prediction in Bangladesh Haor Wetlands

May 19, 2026

This study addresses the critical gap in flash flood early warning for the Haor wetlands of Bangladesh and the inability of existing models to adequately represent backwater dynamics. To overcome these challenges, the authors propose a deseasonalized machine learning ensemble approach that leverages Otsu-threshold-based change detection from Sentinel-1 SAR imagery to construct a proxy indicator for upstream Barak River conditions. A weighted ensemble of Random Forest and XGBoost models enables 72-hour probabilistic flood forecasting, integrated with a three-tier alert system and a BRRI-calibrated rice loss assessment module. By explicitly removing temperature-driven seasonal biases that artificially inflate accuracy, the framework delivers actionable lead times of approximately 36 hours. Evaluated against 77 historical flood events, the model achieves 89.6% accuracy, 87.5% recall, and an AUC-ROC of 0.943, with spatial validation showing 84–91% agreement.

0 citationsRead paper

AgriMind: An Ensemble Deep Learning Framework for Multi-Class Plant Disease Classification

May 15, 2026

This study addresses the inefficiency of smallholder farmers in Bangladesh relying on manual identification of diseases in chili, potato, and tomato crops by proposing an ensemble deep learning framework based on ResNet50, EfficientNet-B0, and DenseNet121. Leveraging transfer learning, the approach freezes ImageNet-pretrained backbones and fine-tunes only the classification heads for ten epochs. Model outputs are integrated via simple averaging of softmax probabilities. Evaluated on the PlantVillage dataset, the method achieves an overall accuracy of 99.23%—with perfect classification (100%) for chili and potato diseases and 99.01% for tomato—while maintaining a real-time inference speed of 53 FPS on an NVIDIA T4 GPU. The work demonstrates that model diversity is crucial for performance gains and reveals that over-reliance on the best-performing individual model can degrade ensemble effectiveness.

0 citationsRead paper

Restoring Rhythm: Punctuation Restoration Using Transformer Models for Bangla, a Low-Resource Language

Jul 24, 2025

Low-resource Bangla suffers from poor readability and degraded downstream performance (e.g., in ASR) due to frequent punctuation omission. To address this, we propose the first large-scale, multi-domain punctuation restoration method specifically for Bangla. We construct a high-quality, manually annotated corpus spanning news, literary, and conversational domains, and introduce a data augmentation strategy combining back-translation and controlled noise injection. Our model employs XLM-RoBERTa-large as the encoder within a sequence-labeling framework enhanced with a Conditional Random Field (CRF) decoder to jointly restore periods, commas, question marks, and exclamation marks in an end-to-end manner. On the news test set, it achieves 97.1% accuracy; on real ASR output texts, it maintains 90.2% accuracy—substantially outperforming all baselines. This work establishes the first open-source Bangla punctuation restoration benchmark, including a curated dataset, effective augmentation techniques, and a robust, generalizable model—thereby filling a critical gap in low-resource language punctuation recovery research.

0 citationsRead paper

Recent publications

Latest Papers

Explainable AI for Screening Abuse-Related Trauma in Bangladeshi Children: A Training-Free Multimodal Framework Evaluated on Noise-Aware Synthetic Data

Jul 04, 2026

This study addresses the critical gap in Bangladesh’s child mental health landscape: the absence of localized, interpretable tools for early psychological trauma screening and severe shortages of specialized mental health resources. To tackle this, the authors propose ShishuRaksha—the first training-free, multimodal AI decision support framework designed for low-resource settings. It integrates four modalities—structured questionnaires, Bengali-language text, House-Tree-Person drawings, and facial emotion cues—employing clinically weighted fusion and cross-modal attention mechanisms to generate bilingual, interpretable reports linked directly to the national child protection system. Key innovations include single-modality coverage rules, perturbation-based additive attribution explanations, tree-ensemble surrogate modeling, and a noise-aware synthetic data benchmark. Evaluated on 500 synthetic cases, the model achieves an AUC of 0.874, significantly outperforming the Strengths and Difficulties Questionnaire (SDQ) unimodal baseline (AUC = 0.756), with ablation and calibration analyses confirming its robustness.

0 citationsRead paper

HaorFloodAlert: Deseasonalized ML Ensemble for 72-Hour Flood Prediction in Bangladesh Haor Wetlands

May 19, 2026

This study addresses the critical gap in flash flood early warning for the Haor wetlands of Bangladesh and the inability of existing models to adequately represent backwater dynamics. To overcome these challenges, the authors propose a deseasonalized machine learning ensemble approach that leverages Otsu-threshold-based change detection from Sentinel-1 SAR imagery to construct a proxy indicator for upstream Barak River conditions. A weighted ensemble of Random Forest and XGBoost models enables 72-hour probabilistic flood forecasting, integrated with a three-tier alert system and a BRRI-calibrated rice loss assessment module. By explicitly removing temperature-driven seasonal biases that artificially inflate accuracy, the framework delivers actionable lead times of approximately 36 hours. Evaluated against 77 historical flood events, the model achieves 89.6% accuracy, 87.5% recall, and an AUC-ROC of 0.943, with spatial validation showing 84–91% agreement.

0 citationsRead paper

AgriMind: An Ensemble Deep Learning Framework for Multi-Class Plant Disease Classification

May 15, 2026

This study addresses the inefficiency of smallholder farmers in Bangladesh relying on manual identification of diseases in chili, potato, and tomato crops by proposing an ensemble deep learning framework based on ResNet50, EfficientNet-B0, and DenseNet121. Leveraging transfer learning, the approach freezes ImageNet-pretrained backbones and fine-tunes only the classification heads for ten epochs. Model outputs are integrated via simple averaging of softmax probabilities. Evaluated on the PlantVillage dataset, the method achieves an overall accuracy of 99.23%—with perfect classification (100%) for chili and potato diseases and 99.01% for tomato—while maintaining a real-time inference speed of 53 FPS on an NVIDIA T4 GPU. The work demonstrates that model diversity is crucial for performance gains and reveals that over-reliance on the best-performing individual model can degrade ensemble effectiveness.

0 citationsRead paper

Restoring Rhythm: Punctuation Restoration Using Transformer Models for Bangla, a Low-Resource Language

Jul 24, 2025

Low-resource Bangla suffers from poor readability and degraded downstream performance (e.g., in ASR) due to frequent punctuation omission. To address this, we propose the first large-scale, multi-domain punctuation restoration method specifically for Bangla. We construct a high-quality, manually annotated corpus spanning news, literary, and conversational domains, and introduce a data augmentation strategy combining back-translation and controlled noise injection. Our model employs XLM-RoBERTa-large as the encoder within a sequence-labeling framework enhanced with a Conditional Random Field (CRF) decoder to jointly restore periods, commas, question marks, and exclamation marks in an end-to-end manner. On the news test set, it achieves 97.1% accuracy; on real ASR output texts, it maintains 90.2% accuracy—substantially outperforming all baselines. This work establishes the first open-source Bangla punctuation restoration benchmark, including a curated dataset, effective augmentation techniques, and a robust, generalizable model—thereby filling a critical gap in low-resource language punctuation recovery research.

0 citationsRead paper