Dr. Mohammad Masum

BiLens combines Vision Transformers and LLM-generated captions to improve product recommendation accuracy. Visual features are extracted using ViT, while image captions are transformed into contextual word embeddings. These representations are integrated through early fusion and compared using cosine similarity to retrieve products that better reflect customer intent.

VisioRAG extends multimodal recommendation by using Retrieval-Augmented Generation and LLM-based caption enhancement. The framework uses Florence-2-large for image captioning, BERT for contextual embeddings, and Gemini for caption enhancement, enabling richer product representations and more human-like query understanding.

FACTOR addresses fine-grained attribute recognition in fashion and e-commerce by improving structured attribute prediction while preserving the stability of pretrained vision-language models. The framework focuses on attributes such as color, material, and pattern, which are essential for interpretable product understanding.

CARVE integrates Grounding DINO, SAM2, and CLIP-based semantic verification to support open-vocabulary product segmentation without task-specific retraining. The framework enables robust product discovery in real-world visual media and significantly improves segmentation quality and recall compared to closed-set baselines.

This project investigates agentic segmentation frameworks that use Gemini 2.5 as a reasoning orchestrator coordinating specialized tools such as Grounding DINO and SAM2. The hybrid framework achieves strong segmentation quality while supporting flexible open-vocabulary reasoning.

This work provides a quantitative comparison between multimodal LLMs and specialized object detection models such as YOLO and DETR. The study evaluates prompting strategies, RAG, and LoRA fine-tuning to understand how general-purpose multimodal models compete in vision-centric tasks.

A diffusion-based inpainting framework designed to generate large-scale synthetic segmentation datasets using prompt-guided scene generation. The framework reduces dependence on manual annotation and improves segmentation performance in data-scarce settings.

HybridSense-LLM is a structured multimodal framework that integrates wearable sensor signals, statistical feature extraction, and prompt-driven LLM reasoning for wellness prediction. The system evaluates multiple prompting strategies across OpenAI, Gemini, and DeepSeek models to estimate stress, fatigue, readiness, and sleep quality from wearable data.

This line of work explores how large language models can reason over structured physiological descriptors and waveform context to support transparent and interpretable health prediction. The research studies the interaction between statistical health features and natural-language reasoning strategies.

The framework systematically evaluates zero-shot, few-shot, self-consistency, and tree-of-thought prompting approaches for health-related prediction tasks. Results show that different prompting strategies provide task-dependent improvements across wellness indicators.

PE2Prompt converts high-dimensional PE malware features into structured language-based prompts for large language model reasoning. The framework uses LLM-guided feature curation, numerical feature summarization, and structured prompt construction to support interpretable zero-shot and few-shot malware classification across models such as GPT-4o-mini, Claude, and DeepSeek.

This extension decomposes PE malware analysis into specialized agents aligned with EMBER feature groups, including sections, linkage, metadata, histogram, byte entropy, and printable strings. Each agent produces an independent verdict and evidence-grounded explanation, which are aggregated through weighted voting and summarized into a final analyst-facing rationale.

This line of work formalizes a new feature-to-language paradigm for cybersecurity, where numerical and structural malware indicators are transformed into natural language reasoning templates. The goal is to make static malware workflows more explainable, adaptable, and easier to audit in real-world security settings.