NCCI 2025 - National Conference on Computer Innovations
Theme: Exploring the Future of Technology and Innovation
The National Conference on Computer Innovations (NCCI) is a premier academic event that brings together researchers, industry professionals, and students to share knowledge, present research findings, and discuss emerging trends in computer science and technology. Hosted by Kathmandu University Computer Club (KUCC) in collaboration with the Department of Computer Science and Engineering (DoCSE), NCCI 2025 aims to foster innovation, collaboration, and knowledge exchange in the rapidly evolving field of computer science. The conference provides a platform for participants to network with peers, engage with industry leaders, and gain insights into cutting-edge research and technological advancements.
Papers Repository
Published papers from NCCI 2025
A Data-Driven Decision Support System for Crop Selection Using IoT Sensors and Machine Learning
by Jatin Bhusal, Jeevan Khatri, Prashanta Acharya, Sanket Shrestha, Ankit Mahato
Full PaperUnpredictable weather patterns, soil degradation, and the unregulated use of pesticides and chemical fertilizers have severely impacted agricultural productivity in Nepal. These challenges have made crop planning increasingly difficult, leading to substantial financial losses and deteriorating mental health among farmers. In response, this study proposes a decision support system that integrates IoT-based environmental sensing with machine learning to provide real-time crop recommendations. The system collects key parameters, including soil pH, NPK levels, moisture content, humidity, soil temperature, and rainfall, which are fed into a supervised learning model to enable precise, data-driven crop selection. Among the models evaluated, the Support Vector Machine (SVM) classifier achieved a prediction accuracy of 96.75% with minimal overfitting, demonstrating the potential of the proposed approach to improve agricultural decision-making.
View Paper →Improved Multimodal Integration with Attention for Live Nepali Sign Language Interpretation
by Ashish Panday
Short PaperAbstract—This work introduces a framework for attention-based multimodal integration in real-time Nepali Sign Language (NSL) interpretation by combining hand landmarks, face details, and pose information. We have fine-tuned on frame series from five NSL gestures as a result the system attains 97.39% validation precision and 31.18 FPS on a GPU. Experiments reveal 83.48% precision in five-epoch scratch training and 46.96% absent ImageNet initialization. The codebase is shared openly to promote NSL inclusivity.
View Paper →Evaluating Sentence Embedding Models for Nepali Sentiment Analysis: A Comparative Study
by Abiral Adhikari, Samir Wagle, Reewaj Khanal, Prashant Manandhar
Short PaperSentiment analysis for morphologically complex, low-resource languages like Nepali remains a developing field, where progress has been largely constrained by a reliance on traditional feature engineering and first-generation neural embeddings. This study confronts this methodological gap through a rigorous comparative analysis designed to decouple the influence of modern embedding representations from downstream architectural complexity. We benchmark four state-of-the-art multilingual sentence embeddings (BGE-m3, LaBSE, E5-base, and DistilUSE) across three neural architectures of increasing complexity: a Multi-Layer Perceptron (MLP), a Residual MLP, and a Transformer network. BGE_m3, when paired with a simple MLP, achieved a notable accuracy of 82.49%, decisively outperforming the more complex Transformer-based classifiers. This result demonstrates conclusively that for this low-resource paradigm, the semantic richness of the input embedding is the dominant determinant of performance, eclipsing the architectural inductive biases of the downstream model. Our work not only establishes a powerful and resource-efficient benchmark for Nepali NLP but also provides crucial insight for sentiment analysis in other low-resource languages.
View Paper →Cognify: Enhancement of Mental Conditions Using Cognitive Tools
by Aarya Pathak, Sarbesh KC, Swoham Kayastha, Aakriti Pandey, Lekhnath S Pathak
Short PaperMental health conditions such as anxiety, depression, and attention-deficit/hyperactivity disorder (ADHD) are increas- ingly prevalent, yet scientifically grounded and accessible tools for their management remain limited. This paper introduces Cognify, an integrative solution combining psychiatry, natural language processing (NLP), and cognitive science to provide a holistic approach to mental health support. The platform initiates with validated psychiatric assessments and continuously adapts through dynamic journal analysis and personalized cognitive interventions. Through sentiment analysis and cognitive task performance, Cognify delivers feedback-driven, evidence-based interventions. By bridging clinical psychiatry with NLP and cognitive tools, Cognify presents a scalable and personalized mental health support system suitable for both clinical and non- clinical contexts.
View Paper →Federated Learning Framework for Scalable AI in Heterogeneous HPC and Cloud Environments
by Sangam Ghimire, Nirjal Bhurtel, Bishal Neupane, Bigyan Byanju Shrestha, Subarna Bhattarai, Prajwal Gaire, Jessica Thapa, Sudan Jha, Paribartan Timalsina
Full PaperAs the demand grows for scalable and privacy-aware AI systems, Federated Learning (FL) has emerged as a promising solution, allowing decentralized model training without moving raw data. At the same time, the combination of high-performance computing (HPC) and cloud infrastructure offers vast computing power but introduces new complexities,especially when dealing with heterogeneous hardware, communication limits, and non- uniform data. In this work, we present a federated learning framework built to run efficiently across mixed HPC and cloud environments. Our system addresses key challenges such as system heterogeneity, communication overhead, and resource scheduling, while maintaining model accuracy and data privacy. Through experiments on a hybrid testbed, we demonstrate strong performance in terms of scalability, fault tolerance, and convergence, even under non Independent and Identically Distributed (IID) data distributions and varied hardware. These results highlight the potential of federated learning as a practical approach to building scalable Artificial Intelligence (AI) systems in modern, distributed computing settings
View Paper →Touch, Swipe, Share: Designing Intuitive Mobile News Interfaces for Gen-Z Engagement in Nepal
by Yunidh Rawal, Prayash Shakya, Mukul Aryal, Anupama Neupane
Full PaperThe rise of short-form content platforms has re-shaped how Generation Z consumes digital media, favoring fast, visual, and socially shareable content. This study explores how intuitive mobile interface design can enhance news engagement among Gen-Z users in Nepal by leveraging familiar interaction patterns such as swipe gestures, touch navigation, and integrated social sharing. We designed a prototype news application inspired by platforms like TikTok and Instagram Reels, and evaluated it using a mixed-method approach: an initial survey (N=110) assessing news habits, and usability testing (N=15) with Gen-Z participants. Results revealed that 75.5% of users rely on social media for news, with 70% preferring short-form delivery formats. Our prototype achieved a System Usability Scale (SUS) score of 82.16 and was rated highly engaging by 86% of participants. These findings demonstrate that gesture-based, mobile-first news interfaces can significantly improve user engagement among younger audiences, offering actionable design guidelines for the future of digital journalism.
View Paper →Sambodhan: Addressing Stigma and Accessibility Through a Youth-Centered Digital Mental Health Solution
by Pratikshya Sapkota, Samish Shrestha
Full PaperAddressing Stigma and Accessibility Through a Youth-Centered Digital Mental Health Solution
View Paper →Braille Voice: A Digital Braille-to-Speech System for Nepali and English Language Accessibility
by Sandip Bajagain, Bibek Timilsina, Samir Adhikari, Niraj Bajagain, Sangram Thapa
Short PaperAccess to information and education is among the foremost challenges confronted by the visually impaired, particularly in under-resourced linguistic settings such as Nepali. Though digital Braille OCR systems seem to achieve more for widely spoken languages, in nations like Nepal, users of Nepali and English Braille have an irreparable accessibility gap. This paper presents Braille Voice-A digital assistive Braille-to-speech multi-platform system for both Nepali and English Braille scripts. Our system uses a purely deterministic classical image processing pipeline under OpenCV with a position-based mapping schema allowing real-time Braille recognition and speech synthesis. Through an iterative design and user-centric testing process, Braille Voice demonstrates robust performance with digitally generated Braille images, while real-world captures expose clear limitations related to environmental factors. This work outlines the project's factual progress, system architecture, and a practical roadmap for dataset creation and advanced processing, aiming to set a rigorous foundation for closing the accessibility divide for visually impaired Nepali and English readers.
View Paper →Acoustic Event Detection and Classification for Monitoring Illegal Logging and Deforestation Using Edge AI Devices
by Ayush Dangol, Ankit Mahato, Majil Budathoki, Shankar Tamang
Full PaperRapid detection of illegal logging is essential for safeguarding the world's shrinking forests. This paper presents a ready-to-publish study on an acoustic monitoring system that employs the pre-trained YamNet model for classifying environmental sounds indicative of deforestation threats. Audio data were manually curated from YouTube and open repositories, converted to log-mel spectrograms, and fine-tuned via transfer learning. A solar-powered edge device streams the classified events to a cloud back-end for real-time alerting. Experimental results on an 18-class forest-sound dataset show an overall accuracy of 87% and F1-scores of 1.0 for chainsaw, gunshot, and hand-saw events. The approach outperforms custom CNN and LSTM baselines by 5%–8% while requiring 40% fewer parameters, demonstrating its suitability for low-power field deployment.
View Paper →Enhancing Ethical Reasoning in Tiny Language Models via Fine-Tuning and Multi-Agent Consensus
by Sushan Adhikari, Sunidhi Sharma, Darshan Lamichhane, Sanjog Sigdel
Short PaperIntegrating ethical reasoning capacity into artificial intelligence systems is a fundamental challenge in the development of trustworthy AI. Although large language models have shown promise in moral reasoning, their computational demand restricts everyday use. In this work, we investigate the possibility of Tiny Language Models (TinyLLMs) for sophisticated ethical reasoning through strategic fine-tuning techniques. We present a framework for adapting TinyLlama-1.1B models through Low-Rank Adaptation(LoRA) on a synthetic dataset of 1,000 ethical dilemmas created using Gemini 2.5 Pro. Our solution develops three expert agents for utilitarian, deontological, and virtue ethics perspectives. The agents use a confidence-weighted consensus model for group decision-making. Qualitative analysis demonstrates considerable improvement in the quality of reasoning: the specialized agents achieved high philosophical consistency, with the deontological, utilitarian, and virtue ethics agents scoring 97.8, 95.2, and 96.5, respectively. Systematic vocabulary analysis also confirms clear differentiation in reasoning styles between the ethical theories. This work verifies that parameter efficient fine-tuning achieves compact models to perform sophisticated moral reasoning suitable for resource-scarce environments.
View Paper →Multi-Stage Fine-Tuning of mT5-Small for Nepali News Summarization
by Akarshan Sekhar Shrestha, Albina Shrestha, Arnav Bhatta, Yunika Bhochhibhoya, Sophiya Shrestha, Prakash Poudyal
Short PaperIn this work, we fine-tune the multilingual text-to-text transformer model, mT5-small, to perform abstractive, 2-3 sentence summarization of Nepali news articles. The model was trained on a custom dataset, which we collected extensively from Nepali news websites. Evaluation using ROUGE metrics yielded the scores ROUGE-1: 46.74, ROUGE-2: 33.01, ROUGE-L: 41.96 and evaluation using BERTScore-F1 gave the score 80.14, demonstrating viable performance for a low-resource language.
View Paper →TikhoFormer: A Two-Stage Blur Classification and Transformer-Based Deblurring Framework
by Mukul Aryal, Suyog Ghimire, Shreyash Poudel, Bigya Vijay Dhungana
Full PaperBlind image deblurring, where the blur type is unknown, remains a significant challenge. State-of-the-art models often employ large, monolithic networks that are computationally expensive. To address this, we propose TikhoFormer, a novel two-stage framework that first identifies the blur type and then applies a specialized deblurring network. The first stage is a highly efficient classifier operating on a compact set of 10 discriminative, hand-engineered features designed to capture blur characteristics. This classifier, trained with Focal Loss and advanced scheduling, achieves 92.5% accuracy and a 0.98 ROC AUC. The second stage deploys one of two lightweight a hybrid CNN-Transformer architecture featuring a U-Net-style encoderdecoder, a Transformer bottleneck for global context, and a dedicated edge-refinement module. Each specialized network is optimized for a single blur type using a composite sharpnessoriented loss. Our complete pipeline, with a total deployable size of only 1.33M parameters, achieves a max PSNR of 27.7 dB and an excellent SSIM of 0.98. This work provides a detailed architectural breakdown and training methodology for both the feature-based classifier and the deblurring network, presenting a complete blueprint for an efficient, expert-guided deblurring system.
View Paper →Developing Ran2Dev: A Model Converting Ranjana Lipi to Devanagari Script (Without Modifiers)
by Nischal Baidar, Sarif Tachamo, Nekesh Koju, Shree Ram Khaitu, Surag Basukala
Short PaperNepal Bhasa (Newar language), traditionally spoken by the Newar community in the Kathmandu Valley, carries significant cultural and historical importance. Over the years, the number of native speakers has gradually declined. In 2024, Nepal Bhasa gained official recognition as an indigenous language of Nepal and currently holds official status in Bagmati Province and local governments such as Kathmandu. One of the primary writing systems for this language is the Ranjana script, which includes 16 vowels, 36 consonants, and 10 numerals. This paper introduces an Optical Character Recognition (OCR) system that converts characters from the Ranjana script into the Devanagari script. To achieve this, two convolutional neural network (CNN) models were used: the standard LeNet-5 and a proposed model named Ran2Dev. The system incorporates processes such as image preprocessing, feature extraction, and model evaluation to enhance accuracy. The Ran2Dev model outperformed LeNet- 5 with an accuracy of 99.74% compared to 99.10%, showing strong effectiveness in recognizing Ranjana script characters.
View Paper →Digital Friction in Public Services: A UX Assessment of the DOTM License Platform in Nepal
by Anupama Neupane
Full PaperWhile e-governance platforms are intended to improve access to public services, many government websites continue to fall short in terms of usability. The Department of Transport Management (DOTM) portal in Nepal, which is used for driving license applications and vehicle registration, reflects several of these challenges. This study looked at the DOTM portal using Jakob Nielsen’s heuristics and feedback from 119 users over two weeks. The main problems were a cluttered layout, confusing navigation, poor system feedback, and bad mobile performance. Users also mentioned broken links, and CAPTCHA issues that made them start over. Based on both the users’ feedback and the heuristic evaluation, five key usability problems emerged, lack of simplicity and visual clarity, limited access to help or documentation, design inconsistencies across pages, difficulty recognizing navigation elements, and unclear system status. These issues not only complicate the user experience but also make it harder to complete essential tasks. Accessibility was another major concern, with screen reader support largely non-functional. To address these problems, the study proposes a redesigned interface with clearer visual structure, consistent navigation, and more helpful system responses. Although the focus is on DOTM, the findings may reflect broader patterns across public digital platforms and highlight the need for more user-centered design in government services.
View Paper →Comparing OCR Engines for Nepal Lipi Extraction
by Pragyan Shrestha, Samriddha Lal Shrestha, Pratik Sharma, Ranjita Dhakal, Prabal Lamichhane, Rajani Chulyadyo
Full PaperNepal Lipi, previously known as Prachalit Lipi, was the native script of Kathmandu and old Nepal before the modern Devanagari script. There are many writings in this script still present to this date in old scriptures, documents, and paintings that hold cultural and historical significance about music, society, and Ayurveda. Optical Character Recognition (OCR) is an important tool in digitization of such artifacts. In this paper, we present a comparison of four OCR engines - EasyOCR, PaddleOCR, TesseractOCR, and CalamariOCR - for recognizing texts written in Nepal Lipi. For this, a custom dataset of 4,625 line segmented images of handwritten manuscripts was prepared for training and testing the models. Of the 4 models, CalamariOCR and PaddleOCR emerged as the top performers with an average Character Error Rate (CER) of 3.755 percent and 9.06 percent, respectively. The trained models were then incorporated into a user-friendly web-app making it a practical tool with enhanced accessibility that can be used for Nepal Lipi’s preservation and opens a gate into further work in this field.
View Paper →Evaluating Lightweight Machine Learning Models for Botnet Detection in IOT: A Performance and Efficiency Perspective
by PRASHANTA ACHARYA, Udit Kumar Mahato
Short PaperThe rapid proliferation of Internet of Things (IoT) devices has expanded the digital landscape, but it has also increased vulnerabilities to botnet-driven cyber threats such as Distributed Denial-of-Service (DDoS) attacks. Traditional security solutions are often too resource-intensive for the limited computational capacity of IoT devices. This study presents a comprehensive evaluation of lightweight machine learning models for botnet detection in IoT environments. Classical algorithms including Naive Bayes, Logistic Regression, Decision Trees, Random Forests, and XGBoost are compared alongside a shallow Convolutional Neural Network (CNN), with a focus on accuracy, interpretability, and resource efficiency. Using a real-world IoT botnet dataset, the models are evaluated based on precision, recall, F1-score, training time, and deployment feasibility in constrained settings. Results show that ensemble methods and lightweight CNNs offer superior detection performance, while classical models like Logistic Regression and Naive Bayes excel in low-latency and energy-limited environments. The findings advocate for adaptive, tiered defense mechanisms tailored to the diverse constraints of IoT deployments.
View Paper →Factors Influencing Nepali Students' Study Abroad Decisions: An Integration of Discrete Choice and Predictive Models
by Mala Deep Upadhaya, Barsha Thakuri
Full PaperGlobal migration for study purposes has been a growing trend, with Nepali students aged 16 and above moving abroad for study purposes. A study examining the socio-economic factors motivating Nepali students (n=101) revealed that factors such as the target study country, presence of family abroad, and part-time work opportunities were significant. In contrast, gender, age, and education level showed minimal impact. Lower living costs in environmentally diverse regions and university prestige were also prime factors. The research developed an abroad study prediction model using six machine algorithms and various sampling techniques like random undersampling, random oversampling, SMOTE, and ADASYN, to address class imbalances and prevent data leakage. The study found that logistic regression with RandomOverSample followed by SMOTE performed better than other methods, with best mean accuracy of 0.83 and 0.82, respectively. The study also used the ethical approach of the right to delete data promptly and a clean slate for socio-demographic analysis. However, the study acknowledges limitations, including a small sample size, potential influences from psychosocial factors, and the cross-sectional nature of the data. Future research with larger samples and longitudinal studies could provide deeper insights into the complex decision-making processes of students.
View Paper →A Comparative Study of Machine Learning Algorithms for Document Classification: Insights Beyond Accuracy
by Ankit Mahato, Udit Kumar Mahato
Full PaperDocument classification remains a fundamental task in natural language processing with applications across diverse domains. While state-of-the-art transformer-based models like BERT demonstrate superior accuracy, their computational and environmental costs often outweigh marginal performance gains. This study provides a comprehensive comparative analysis of classical machine learning models (SVM, Naïve Bayes, Logistic Regression), deep learning architectures (CNN, RNN), and transformer-based models (BERT, DistilBERT, LayoutLM). We evaluate models across multiple dimensions including accuracy, training efficiency, energy consumption, robustness, and interpretability. Our results highlight the continued relevance of classical models in resource-constrained environments and introduce an Energy-Adjusted Score (EAS) to balance performance and cost. We also expose issues in benchmark datasets that challenge conventional performance reporting. Based on empirical insights, we offer algorithm selection guidelines tailored for specific deployment scenarios.
View Paper →An Artificial Intelligence Driven Semantic Similarity-Based Pipeline for Rapid Literature Review
by Abhiyan Dhakal, Kausik Paudel, Sanjog Sigdel
Full PaperWe propose an automated pipeline for performing literature reviews using semantic similarity. Unlike traditional systematic review systems or optimization-based methods, this work emphasizes minimal overhead and high relevance by using transformer-based embeddings and cosine similarity. By providing a paper’s title and abstract, it generates relevant keywords, fetches relevant papers from open-access repositories (e.g., ArXiv), and ranks them based on their semantic closeness to the input. Three embedding models: TF-IDF, all-MiniLM-L6-v2, and Specter2 were evaluated. While TF-IDF struggled with capturing deeper semantic meaning, all-MiniLM-L6-v2 provided broader conceptual coverage. Specter2, specifically fine-tuned for scientific texts, exhibited score saturation in similarity scores. A statistical thresholding approach is then applied to filter relevant papers, enabling an effective literature review pipeline. Despite the absence of heuristic feedback or ground-truth relevance labels, the proposed system shows promise as a scalable and practical tool for preliminary research and exploratory analysis.
View Paper →Detecting Image Forgeries and Deepfakes: A Comparative Study of CNN and Transformer Models with a Custom-Curated Dataset
by Sushan Adhikari, Pranish Kafle, Aayush Man Shakya, Nitin Ghimire, Gajendra Sharma
Full PaperThe exponential growth of sophisticated image manipulation technologies, including deepfakes and traditional forgery techniques, has created an urgent demand for robust detection mechanisms in digital forensics and computer vision. This research investigates the comparative effectiveness of two distinct deep learning architectures—InceptionV3 and Vision Transformer (ViT)—for identifying manipulated and synthetic imagery. Our experimental methodology utilized a comprehensive dataset containing over 140,000 images, encompassing conventional manipulation methods such as photoshopping, splicing, copy-move operations, and face-swapping, alongside AI-generated content from StyleGAN, StyleGAN2, and deepfake technologies. Through rigorous evaluation protocols, the InceptionV3 model demonstrated superior performance with 94.0% test accuracy and 93.96% validation accuracy, while the Vision Transformer achieved 88.49% test accuracy. Comprehensive performance analysis across multiple evaluation metrics—including precision, recall, F1-score, and computational efficiency—revealed that InceptionV3 outperformed ViT by 5.51% in accuracy and 10.8% in recall performance. These findings challenge prevailing assumptions about transformer architectures universally surpassing CNNs in computer vision applications. The results indicate that CNN-based architectures, particularly InceptionV3, provide substantial advantages for image forensics applications through enhanced computational efficiency and superior detection capabilities for both manipulated and synthetically generated content. This research contributes valuable insights into architectural selection for deepfake detection systems and establishes benchmarks for future developments in digital media authenticity verification technologies.
View Paper →eBPF-PATROL: Protective Agent for Threat Recognition and Overreach Limitation using eBPF in Containerized and Virtualized Environments
by Sangam Ghimire, Roshan Sahani, Sudan Jha, Nirjal Bhurtel
Full PaperWith the increasing use and adoption of cloud and cloud-native computing, the underlying technologies,(i.e containerization and virtualization) have become foundational. However, strict isolation and maintaining runtime security in those environments has become increasingly challenging. Existing approaches like seccomp and Mandatory Access Control (MAC) frameworks offer some protection upto a limit, but often lacks context-awareness, syscall argument filtering, and adaptive enforcement. Our paper introduces eBPF-PATROL (eBPF-Protective Agent for Threat Recognition and Overreach Limitation), an extensible lightweight runtime security agent that uses extended Berkeley Packet Filter (eBPF) technology to monitor and enforce policies in containerized and virtualized environments. By intercepting system calls, analyzing execution context, and applying user- defined rules, eBPF-Protective Agent for Threat Recognition and Overreach Limitation (eBPF-PATROL) detects and prevents real-time boundary violations, such as reverse shells, privilege escalation, and container escape attempts. We describe the architecture, implementation, and evaluation of eBPF-PATROL, demonstrating its low overhead (<2.5%) and high detection accuracy across real-world attack scenarios.
View Paper →Multi-Layered Cyber Defense: Combining AI-Based Malware Classification, Server Monitoring, and Penetration Testing
by Rupeshkhadka4, Prajwal Rai , Bibek Gautam
Full PaperA multi-layered cybersecurity framework is suggested to tackle the growing complexity of cyber threats that frequently bypass traditional signature based antivirus solutions. The system combines three main elements: AI-based malware classification, live server monitoring, and penetration testing, all developed with the Flask framework and presented via a user friendly web interface. Highlighting flexibility and quick responsiveness, the system utilizes two machine learning algorithms Extra Trees Classifier and Logistic Regression for malware detection and classification, with the Extra Trees model reaching an outstanding accuracy of 99.24%. This elevated precision highlights its ability to detect both familiar and new threats. The server monitoring component identifies irregularities like unauthorized access attempts and unusual resource usage, whereas the penetration testing feature uncovers system weaknesses using controlled attack simulations. These elements are linked in a dynamic feedback loop, enabling the system to adjust instantly to new threats by exchanging information across layers. This closed-loop system greatly enhances threat detection accuracy, response effectiveness, and overall system resilience. Through continuous monitoring, smart classification, and proactive risk evaluation, the system provides an all-surround and flexible approach to safeguarding vital infrastructure and business settings. Its flexible and expandable design renders it appropriate for deployment in various industries, including developed as well as developing areas where cybersecurity resources may be uneasy. The research shows that combining machine learning with ongoing monitoring and automated testing improves security defenses and validity against the changing nature of cyber threats in the modern digital environment.
View Paper →