The formal record

Independent Research · BRAC University

• ▸Designed Obscrowd, a privacy framework for multi-portrait biometric unlinkability under generative editing, protecting every detected face in a group image with a single imperceptible, mask-guided perturbation.
• ▸Engineered a teacher-student-refinement pipeline: a diffusion-guided teacher, a lightweight student generator for efficient deployment, and an inference-time refinement stage within a bounded perceptual budget.
• ▸Reduced face-recognition similarity from 0.957 to 0.138 after generative edits while keeping the protected image visually faithful, training against a face-recognition ensemble and differentiable editing attacks.

Brain Tumor Segmentation & Classification

• ▸Implemented and compared U-Net, Attention U-Net, and three CNN classifiers (DenseNet-121, EfficientNet-B0, MobileNetV2) on the BRISC 2025 MRI dataset in PyTorch.
• ▸Achieved 88.22% test Dice, 79.74% mIoU, and 99.61% pixel accuracy for segmentation, and 97.50% four-class classification accuracy on 860 unseen samples.
• ▸Ran a 20-configuration hyperparameter study and demonstrated that separate task-specific training outperforms joint multi-task learning by ~5 points, informing the deployment recommendation.

News Classification & AI-Text Detection

• ▸Built a reproducible benchmark of 27 experiments (9 architectures x 3 preprocessing pipelines) on 102k news headlines, reaching 0.9376 macro-F1 with BERT-Base and 0.9214 with a Bi-GRU trained in 19 seconds.
• ▸Engineered 28 stylometric features with mRMR and RFE selection for AI-generated text detection, reaching 97.40% accuracy on HC3 with Random Forest and quantifying a severe cross-generator generalization gap.
• ▸Identified that BERT's WordPiece tokenizer inverts the preprocessing rules that hold for classical models, a model-specific insight for practitioners.

BRACU Vault · Polaris · E-GamerHub

• ▸Built BRACU Vault, a verified study-archive platform, on Cloudflare's free tier (Workers + Hono, D1, R2, Clerk) with a data-driven six-role RBAC model and university-email-gated uploads.
• ▸Engineered Polaris, an AI academic strategist using Gemini structured-JSON output, RAG over a curated knowledge base, and a transparent logistic-regression probability model, with graceful no-API-key fallback.
• ▸Developed E-GamerHub, a MERN social platform with Socket.IO real-time messaging, RBAC moderation, and Groq Llama 3 natural-language player search.