I am a PhD student and Graduate Research Assistant at Louisiana State University (LSU), where my research sits at the intersection of computer vision, large language models, and multimodal AI. My current doctoral work focuses on privacy-preserving face-swapping pipelines — transferring facial identity while formally protecting sensitive biometric attributes — directly contributing to Trustworthy AI and data privacy research.
In parallel, I am the founder of Gradmate.ai — an AI-powered platform streamlining graduate school applications through LLM orchestration, semantic program matching, SOP generation, and IELTS practice. Prior to LSU, I was a Graduate Research Assistant at Alfred University, where I won 1st place (tied) in the Advizex AI Innovation Challenge, developed an AI sales bot projected to increase ROI by 25%.
My publication record spans IEEE conferences, ICPR, and arXiv, with work in facial attribute captioning, sign-language recognition, low-resolution license plate recognition, and toxic content classification. I also have open datasets and fine-tuned models on HuggingFace actively used by the community.
Current PhD research developing a face-swapping pipeline that transfers facial identity while formally protecting sensitive biometric attributes from inference attacks. Directly contributing to Trustworthy AI and data privacy, with techniques for identity disentanglement in generative models enabling controlled face synthesis with formal privacy guarantees.
A face-swapping pipeline disentangling facial identity from expression and appearance in StyleGAN's W-space. Introduces DeepFace neutral-emotion filtering to source neutral-expression StyleGAN faces, paired with InsightFace gender validation for controlled attribute transfer. W-Space Refined blending outperforms Hybrid fusion (ID↔Src cosine similarity 0.359 vs. 0.232), enabled by convex-hull face masks from 106 landmarks, fixed Poisson blending with ROI boundary clamping, and pre-blend colour correction.
FaceGemma enhances a Gemma-based vision-language model with an explicit facial attribute encoder, producing rich, identity-aware captions for portrait images. Trained on the custom FaceAttDB multilingual dataset (Zenodo), the model improves BLEU and CIDEr metrics over general-purpose captioners by conditioning generation on age, gender, emotion, and facial geometry.
Systematic empirical analysis of how LLMs degrade on previously learned NLP tasks after sequential domain fine-tuning. Benchmarks span GLUE tasks (SST-2, MRPC, CoLA, MNLI) across sub-10B open-weight models (Orca, Qwen). Explores elastic weight consolidation, LoRA merging, and replay strategies as mitigation techniques. Identified models with minimal forgetting for robust continual learning. Under review at ACL ARR 2024.
Founded and engineered Gradmate.ai, an end-to-end LLM-powered platform for graduate applicants. Core features include semantic program discovery (dense retrieval over university programme embeddings), multi-agent SOP/LOR generation via LangGraph, IELTS adaptive practice, and a community knowledge base. Stack: LangChain · LangGraph · Groq · DeepSeek · Next.js. Seeking seed investment to scale.
Implemented a three-agent system using Gemma LLMs for Kaggle data tasks: Agent-1 (prompt generation), Agent-2 (context retrieval & elaboration), Agent-3 (suggestion synthesis). Backed by TensorFlow, PyTorch, and JAX. Also developed a reflection-based prompt engineering technique using Llama 3.1 for the LLM 20 Questions competition.
Doctoral thesis developing a multi-agent LLM system integrated with GE Vernova's GridOS platform for real-time power grid decision support. Agents handle fault diagnosis, load-balancing recommendations, and operator-alert summarisation using LangChain tool-calling over ADMS/AEMS APIs. Evaluated against operator response latency and decision-accuracy benchmarks.
NSF-funded summer research applying physics-informed neural networks to predict battery performance metrics including energy density and fast-charging rates. Explored AI-driven material discovery workflows to accelerate next-generation battery development.
Leveraged Phi-3 Vision to analyze thermal and RGB power line imagery, detecting rust, hot spots, and flashover damage for infrastructure safety optimization. Demonstrates multimodal AI applied to critical energy infrastructure inspection workflows.
A two-stage deep-learning pipeline for recognising Bangla license plates under real-world low-resolution conditions. YOLO-based detector localises the plate region, followed by a CNN-LSTM sequence model for character recognition. Evaluated on a custom surveillance dataset achieving state-of-the-art accuracy for Bangla script.
A real-time Bangla Sign Language finger-spelling system built on a custom-trained YOLOv8 object detector operating at 30+ FPS on commodity hardware. Curated dataset of 36 Bangla hand-gesture classes under varied lighting and background conditions, achieving 94.3% mAP. Published at ICRIC 2023.
An automatic portrait colorization framework leveraging identity-aware features extracted from a pretrained VGG-Face network as conditioning signals for a CNN colorization decoder. By anchoring colour prediction to facial identity features, the model produces perceptually plausible skin-tone and hair-colour outputs consistent across different images of the same subject.
- Developing privacy-preserving face-swapping pipeline — Trustworthy AI contribution
- Sequential fine-tuning experiments on GLUE/SuperGLUE benchmarks
- LLM copilot system for GE Vernova GridOS (thesis)
- LangChain & LangGraph for multi-agent LLM orchestration
- Groq and DeepSeek for low-latency LLM inference
- RAG-based program discovery and SOP generation pipelines
- System architecture, model deployment, scalability
Baton Rouge, LA
Alfred, NY
Bangladesh
Bangladesh
Professor of Electrical Engineering
Alfred University
fwangx@alfred.edu · (607) 871-2548
Associate Professor, Material Science & Eng.
Alfred University
wang@alfred.edu · (607) 871-2729
Professor, Dept. of CSE
Ahsanullah University of Science & Technology
kalpoma@aust.edu