Naimul Haque

Face-swapping system transferring identity while formally protecting sensitive biometric attributes from inference attacks, using identity disentanglement in generative models for controlled synthesis with privacy guarantees.

Disentangles identity from expression in StyleGAN's W-space using DeepFace neutral-emotion filtering and InsightFace gender validation. W-Space Refined blending achieves cosine similarity 0.359 vs. 0.232, using convex-hull masks (106 landmarks) and Poisson blending.

Gemma VLM with a facial attribute encoder trained on multilingual FaceAttDB. Conditions generation on age, gender, emotion, and facial geometry, improving BLEU and CIDEr over general captioners.

Empirical study of LLM degradation after sequential fine-tuning across GLUE tasks on sub-10B models (Orca, Qwen). Evaluated EWC, LoRA merging, and replay as mitigations. Under review at ACL ARR 2024.

LLM platform featuring dense-retrieval program discovery, multi-agent SOP/LOR generation via LangGraph, and IELTS adaptive practice. Stack: LangChain · Groq · DeepSeek · Next.js.

Three-agent Gemma pipeline: prompt generation → context retrieval → suggestion synthesis. Built with TF/PyTorch/JAX. Also features a reflection-based Llama 3.1 prompting technique for the LLM 20 Questions competition.

Multi-agent LLM system for real-time power grid decision support: fault diagnosis, load-balancing, and operator alert summarisation via LangChain tool-calling over GE's ADMS/AEMS APIs.

NSF-funded project using physics-informed neural networks to predict battery performance (energy density, fast-charging rates), accelerating next-generation material discovery.

Phi-3 Vision analyses thermal and RGB power line images, detecting rust, hot spots, and flashover damage for critical energy infrastructure inspection.

Two-stage pipeline: YOLO-based plate detector + CNN-LSTM character recogniser. SOTA accuracy on a custom surveillance dataset for Bangla script.

YOLOv8 model for Bangla Sign Language finger-spelling at 30+ FPS. 36-class dataset under varied conditions; 94.3% mAP.

Identity-aware colorization using VGG-Face features to condition a CNN decoder. Produces consistent skin-tone and hair-colour outputs across images of the same subject.