Resume

Obscrowd, BRAC University

• Designed Obscrowd, a multi-portrait biometric privacy framework that protects every detected face with one soft-mask-guided perturbation while preserving the visual fidelity of the original image.
• Built a teacher-student-refinement pipeline combining a diffusion-guided teacher, lightweight student generator, face-recognition ensembles, and differentiable editing attacks for identity disruption.
• Evaluated protection per face across generative edits and processing conditions; a sample smile edit reduced recognition similarity from 0.957 to 0.138, while compression remains an open limitation.

Brain Tumor Segmentation and Classification

• Implemented U-Net and Attention U-Net segmentation plus DenseNet-121, EfficientNet-B0, and MobileNetV2 classification pipelines for four-class MRI analysis on the BRISC 2025 dataset in PyTorch.
• Achieved 88.22% test Dice, 79.74% mIoU, and 99.61% pixel accuracy with U-Net, while DenseNet-121 reached 97.50% classification accuracy on held-out BRISC MRI samples from the four-category test set.
• Ran 20 optimizer and learning-rate configurations and compared shared-encoder multi-task training with separate models, finding task-specific training improved both tasks and avoided task interference.

News Classification and AI-Generated Text Detection

• Built a reproducible 27-run benchmark of nine architectures and three preprocessing pipelines across 102,002 training headlines, spanning TF-IDF, recurrent networks, and BERT-Base under matched controls.
• Reached 0.9376 macro-F1 with BERT-Base and 0.9214 with Bi-GRU trained in 18.7 seconds, showing that aggressive preprocessing helps shallow models but degrades pretrained transformers for this corpus.
• Engineered 28 lexical, syntactic, readability, and distributional features for AI-text detection; Random Forest reached 97.40% in-domain accuracy but exposed severe cross-generator failure.

BRACU Vault and Polaris

• Built BRACU Vault, a verified academic resource platform on Next.js, Cloudflare Workers, Hono, D1, R2, Drizzle, and Clerk, with university-email-gated uploads, moderation, and audit trails.
• Designed a six-role, data-driven access model and controller-service-repository-policy API; atomic D1 activation and audit logging keep verification idempotent and traceable across privileged workflows.
• Built Polaris using Gemini structured JSON, retrieval over a 60-document knowledge base, and a transparent logistic-regression model for explainable academic planning and what-if simulation.
• Implemented offline heuristic fallbacks, Zod validation, subscription gating, MongoDB indexes, signed payment webhooks, and role-aware student, parent, partner, and admin views across paid tiers.

Vantage Arm Digital Twin and WattsUp Office Monitor

• Engineered a Three.js robotic-arm digital twin with damped-least-squares inverse kinematics, forward-kinematics verification, collision checks, joint limits, motion replay, and autonomous PIN entry.
• Centralized manual, voice, Gemini-planned, replay, and PIN commands behind one deterministic safety gateway, with a Zustand state model and an ESP32 six-servo hardware concept for all control modes.
• Built WattsUp as a simulated IoT energy monitor whose Express backend owns 15-device state, virtual time, usage, and alert rules shared by a live Three.js dashboard and Discord bot without client drift.
• Synchronized controls over Socket.IO and exposed status, usage, simulation, turn-off, and alert commands through Discord.js, with stable IDs for deduplicated notifications across both interfaces.

CSE489 Mobile Application Development

• Built two Kotlin Android applications covering Activities, Fragments, lifecycle state restoration, ViewBinding, Jetpack Navigation, responsive XML resources, and JUnit-tested business logic.
• Implemented custom broadcasts, battery status handling, Glide image loading, matrix-based pinch zoom and panning, plus audio and video playback with lifecycle-aware resource cleanup for lifecycle safety.

Custom UNIX Shell and VSFSck

• Built C systems tools: a POSIX shell with fork/exec, pipes, redirection, chaining, history, and scoped SIGINT handling, plus an fsck-style checker for superblocks, bitmaps, inodes, and block references.