Projects
Research work, engineering projects, and creative experiments
Are Biological Foundation Model Benchmarks Actually Trustworthy?
Mar 2026 – PresentSystematically auditing data contamination in protein and RNA language models by adapting membership inference, n-gram overlap, slot guessing to biological sequences. Quantifies SOTA performance benchmarks (Rfam, ProteinGym etc) inflation on train-test overlap. Proposes temporal-split and homology-aware decontamination protocols for reliable future evaluation.
SplatCabin: Privacy-Preserving Synthetic Driver Monitoring via 3D Gaussian Splatting
Mar 2026 – PresentBuilt an end-to-end research pipeline that reconstructs photorealistic car interiors using 3D Gaussian Splatting, composites diverse SMPL-X human avatars via depth-aware rendering, and generates large-scale annotated synthetic datasets for driver monitoring. Trained and evaluated DMS classifiers (ResNet-50, ViT) on synthetic data with successful domain transfer to real-world benchmarks.
Enhanced Dimensional Object Detection in Real-Time
Nov 2024 – Mar 2025Implemented real-time object detection with dimensional measurement using camera calibration and focal-length-based depth estimation. Supported live camera input and reported object classes with confidence scores.
Custom CNN for Retinal Disease Classification
Aug 2025 – PresentDesigned and evaluated custom CNN architectures (6–12 layers, 2.9M parameters) for 4-class retinal fundus image classification, achieving 94.13% test accuracy. Performed ablation studies to analyze architectural components and implemented Grad-CAM for model interpretability.
Bengali Sentiment Classification using Custom LLM
Jun 2024 – Feb 2025A machine learning system for sentiment analysis of Bengali text using custom language models and transfer learning techniques. Developed as undergraduate thesis project.
NLP-Based Bengali Named Entity Recognition
Nov 2023 – Jan 2024A Named Entity Recognition dataset and pipeline for Bengali language using transformers and transfer learning.