Md. Mutasim Billah Abu Noman Akanda

Core methodology: choosing or engineering the evaluation metric and model design that stays honest when the default one misleads — comparative IoU instead of a single detection score, F1 over accuracy under class imbalance, accuracy/FLOPs/latency trade-offs instead of accuracy alone, parse-success-gated scoring instead of raw OCR accuracy.

• Document AI & OCR for low-resource languages: layout analysis and comparative detection evaluation (ACM SE'24 co-first author, Bengali BADLAD); parse-success-gated scoring for handwritten OCR.
• Efficient multimodal adaptation: parameter-efficient fine-tuning, quantization, and sweep-guided model selection for VLMs/LLMs.
• Biomedical image analysis (supporting evidence): diabetic retinopathy grading — unpublished capstone, manuscript in preparation, F1 over accuracy under class imbalance — plus DICOM MRI/CT pipelines on RSNA brain tumor and abdominal trauma tasks.
• Deployable foundation models: reproducible multi-GPU training, objective-aligned metrics, and edge-ready inference.

Diabetic retinopathy detection from fundus photographs (CSE499 capstone, ULAB, examined Mar 2023). First and corresponding author, with Dr. Muhammad Abul Hasan as supervisor and Dr. Mohammed Ashikur Rahman as co-supervisor. Unpublished and never submitted to any venue; a manuscript is in preparation for journal submission.

• Benchmarked custom CNN, ResNet-18/34/101/152, EfficientNet-B0/B1, and ViT-16 (all trained from scratch, no pretraining) on EyePACS (35,126 fundus images, 5-class DR severity); chose F1 over accuracy for class imbalance.
• Proposed architecture RevampResNet-18 (modified ResNet-18: enlarged first-conv kernel/stride, 512×512 input) reached 89% test F1 at 0.88B FLOPs vs. baseline ResNet-18's 83% at 1.79B FLOPs.
• Deployed as DRDetect, a Streamlit/Hugging Face Spaces demo — technical prototype only, not clinically validated, not IRB-approved, not used on patient data; no external validation cohort or comparison against current literature performed.

Code →

Goal: bridge rigorous research and deployable systems for high-variability documents and low-resource scripts.

• Robustness: methods that hold up under layout variance, scan artifacts, and domain shift.
• Evaluation: reproducible benchmarks and ablations that make trade-offs explicit (accuracy vs cost/latency).
• Efficiency: studying accuracy–memory–latency trade-offs when adapting foundation models (quantization, sweep-guided selection, multi-GPU reproducible training).
• Systems: executable research artifacts (APIs, Docker, CI) to reduce “it works on my machine” gaps.