2025 - 2026 Major Project NLP XAI

Distinguishing Human and AI-Generated Nepali Text Using Explainable AI

A hybrid classification framework combining transformer ensembles with SHAP interpretability to detect AI-generated text in low-resource Nepali, addressing misinformation and academic integrity at the language frontier.

Try It

As AI-generated text becomes increasingly indistinguishable from human writing, the challenge of detection is amplified for low-resource languages like Nepali, which lack dedicated corpora and detection tools.

This project developed an ensemble classification framework that fine-tunes multiple transformer architectures, NepBERTa, Multilingual MiniLM, and XLM-RoBERTa, and combines their outputs with traditional ML models to maximize robustness across diverse writing styles and contexts.

Critically, the system integrates SHAP (SHapley Additive exPlanations) at the token level, making the model's decisions transparent and interpretable. This allows end-users, educators, journalists, and policymakers to understand why a text was flagged, not just that it was flagged.

Supervised by Asst. Prof. Bikram Shah · IOE Purwanchal Campus

Stack & Methods

  • Python / PyTorch
  • NepBERTa
  • Multilingual MiniLM
  • XLM-RoBERTa
  • SHAP (XAI)
  • Ensemble Learning
  • Scikit-learn
  • Hugging Face
2024 - 2025 Minor Project Computer Vision

A Lightweight Dual Aggregation Transformer for Image Super-Resolution

A resource-efficient super-resolution architecture using dual spatial and channel transformer blocks with pixel shuffle to reconstruct high-fidelity images at low computational cost.

Try It

Image Super-Resolution (SR) is the task of reconstructing a high-resolution image from a low-resolution input. While recent transformer-based methods have achieved impressive results, most are too computationally expensive to run on edge devices like smartphones or embedded systems.

The Lightweight Dual Aggregation Transformer (LDAT) addresses this gap by introducing two complementary attention blocks: the Dual Spatial Transformer Block (DSTB), which captures spatial feature patterns, and the Dual Channel Transformer Block (DCTB), which models inter-channel dependencies.

A custom Spatial-Guided Feedforward Network (SGFN) further enriches local feature extraction, while pixel shuffle enables efficient upscaling without expensive transposed convolutions, achieving high-resolution reconstruction at a fraction of typical computational cost.

Supervised by Asst. Prof. Pukar Karki · IOE Purwanchal Campus

Stack & Methods

  • Python / PyTorch
  • Transformer Architecture
  • Spatial Attention (DSTB)
  • Channel Attention (DCTB)
  • SGFN
  • Pixel Shuffle
  • Edge Deployment