Ayura – A digital platform providing personalized dietary guidance based on Ayurvedic and Siddha principles in Sri Lanka
Ayurveda and Siddha are traditional medical systems that remain widely used in Sri Lanka. These approaches emphasize natural healing, customized meals, and lifestyle choices based on body type (Vata, Pitta, and Kapha). While these principles are still recognized, many urban residents lack sufficient guidance on how to integrate them into their daily lives.
The primary issue identified is that individuals cannot determine their Ayurvedic body type and apply suitable dietary or lifestyle advice. This lack of personalized treatment leads to generic or inappropriate diets, which can exacerbate chronic conditions like diabetes and digestive issues, as well as stress. Although some traditional healthcare professionals offer advice, access is limited, and the process is not always straightforward for modern users.
This project seeks to bridge that gap by providing a practical and individualized approach to apply Ayurvedic and Siddha concepts. The goal is to help users recognize their body types and obtain meal ideas based on their health situation, aspects such as the season, weather or eating habits.
To collect data, the author conducted literature studies, user surveys, and expert consultations with Ayurvedic doctors. The author also examined existing wellness applications to discover missing functionality. Based on these findings, a prototype called Ayura was created.
The proposed solution is more than just digitizing Ayurveda, it also aims to make old wisdom more accessible. It combines body type detection, seasonal dietary suggestions, and supports users in making better lifestyle choices. This also encourages the ongoing use of Sri Lanka’s diverse traditional healing methods in a present, user-friendly approach.
Sparse-MDI: Adaptive Sparsification for Communication-Efficient Model Distributed Inference in The Edge
The rapid growth of edge computing applications, such as autonomous vehicles and
IoT systems, necessitates real-time, low-latency inference of deep neural networks in
environments with limited computational resources. Most existing distributed inference methods face significant communication overheads when handling large, complex models, especially in resource-constrained edge networks.
This research explores an adaptive sparsification technique by introducing a novel peer-
2-peer (P2P) model distributed inference (MDI) framework to address these communication
challenges. Adjusting the sparsity of activations based on network and device conditions can be adopted by any peer-2-peer MDI framework to improve communication overhead. By dynamically adjusting data loads, adaptive sparsification enhances performance, reduces latency, and preserves model accuracy.
VoiceBlockly: Voice Code Generation in Block-Based Programming Using a Novel Multi-Agent Framework
Block-based programming (BBP) has proven effective in teaching programming concepts, offering a visual and more intuitive approach than traditional text-based programming. However, accessibility in BBP remains an underexplored area, particularly for students with disabilities who may benefit from alternative input methods, such as voice commands. This project aims to bridge this accessibility gap by integrating natural language processing (NLP) capabilities into a BBP interface, enabling users to generate programming blocks through voice commands.
To achieve effective voice to block code generation, the project employs a novel multi-agent (MA) framework that integrates several large language models (LLM) working together. Each model is designed to interpret user commands and generate corresponding block syntax. A novel aggregation algorithm based on uncertainty quantification (UQ) is introduced to combine outputs from multiple agents, ensuring the accuracy and validity of the generated blocks. This approach enables the system to produce high quality output by leveraging the collective strengths of multiple agents.
The prototype demonstrates that the BBP environment effectively translates NL input into accurately rendered code blocks within the interface. Testing and evaluation results indicate that the proposed MA framework outperforms the leading state-of-the-art (SOTA) MA framework by 6% in programming tasks, achieving a Pass@1 score of 75% on the HumanEval dataset. Furthermore, it surpasses the same MA framework by 6% in general domain tasks, attaining an accuracy of 65% on the TruthfulQA dataset.
Machine Learning based Respiratory Sound Analysis for Disease Detection (Pulmo Sense)
Respiratory illnesses like COPD and asthma offer global health implications. Early detection is vital yet problematic due to limited diagnostic technologies and manual assessment by healthcare professionals. This project addresses the need for a machine learning based system to detect respiratory problems through analyzing respiratory sounds.
An ensemble model utilizing feature CNNs was created, including audio preprocessing approaches such as noise handling and segmentation to preprocess breathing sound data for analysis. Each feature CNN consisted of eight layers. Features like Mel-Frequency Cepstral Coefficients (MFCC), Chroma, and Mel-Spectrogram were retrieved to facilitate disease classification. The model was subjected to hyperparameter optimization and cross-validation to enhance its performance.
Preliminary testing shows promising results, with the model achieving an accuracy of 86% and sensitivity of 88% on a diverse dataset of lung sounds. These results indicate the model’s potential in supporting early disease detection, even outside clinical settings.
A Hybrid VAE and GNN-Enhanced Few-Shot Learning Approach for Network Intrusion Detection and Adaptation to Novel Attack Classes
Advanced Network Detection Systems are important to detect established and unknown network attacks, as traditional signature-based methods fail against novel threats and supervised machine learning requires extensive labeled data for new attacks. This study proposes a hybrid deep learning approach: an unsupervised Variational Autoencoder (VAE) for anomaly detection, coupled with a GNN-enhanced Few-Shot Learning (FSL) classifier. The VAE, trained solely on ‘Normal’ data from the UNSW-NB15 dataset, identifies anomalies using a high-percentile reconstruction error threshold. Subsequently, a Prototypical Network (ProtoNet), as the FSL classifier, is episodically trained on ‘Normal’ data and a select subset of previously seen attack types to classify these anomalies. Using these, the malware through network traffic is prioritized for newer variants using the few shot component, after the VAE detects the first stage.
OptiBooking: Hotel Booking Cancellation Prediction and Overbooking Recommendation
OptiBooking is a predictive decision-support system developed to address the challenges of hotel booking cancellations and ineffective overbooking strategies within the Sri Lankan hospitality industry. The solution is specifically designed for small to medium-sized hotels, which form the majority of Sri Lanka’s hotels and are most vulnerable to fluctuations in guest reservations. The system provides a web interface that allows front desk staff, managers, and super users to perform key tasks such as adding new bookings manually or uploading bulk bookings via CSV files, viewing and managing reservations, monitoring cancellation likelihoods, and implementing overbooking strategies. Different user roles are defined to provide access privileges to maintain security and usability across staff levels. The backend of the system is built using Python, where core booking operations, cancellation probability predictions, and automated customer email notification features are handled. A trained machine learning model predicts the likelihood of cancellation based on booking characteristics, supporting proactive operational decision-making. Persistent booking data, room data, and prediction results are stored in a MySQL database. The OptiBooking project not only improves operational efficiency by reducing last-minute cancellations and increasing occupancy rates but also promotes sustainable digital transformation in the hospitality sector. Designed with modularity, accessibility, and future scalability in mind, OptiBooking lays a strong foundation for real-world adoption and continued enhancement.
Towards Explainable and Occlusion Aware Crowd Anomaly Detection
Crowd anomaly detection is a critical research area that addresses the growing need for ensuring safety and security in densely populated urban environments. Traditional CCTV-based surveillance systems often struggle with real-time detection of suspicious activities due to challenges such as occlusions, crowded scenes, and complex human behaviors. VUEBLOX project proposes an advanced, explainable, and occlusion-aware framework for robust crowd anomaly detection. The system integrates multiple deep learning modules, including Masked Autoencoders (MAE) to handle occlusions by reconstructing partially visible objects, Graph Neural Networks (GNN) to capture intricate spatial relationships, and SimCLR for contrastive feature representation. Furthermore, the model employs an ensemble voting mechanism to aggregate outputs from different modules and improve anomaly detection accuracy.
Explainability is a key focus of this framework, achieved through techniques such as heatmap visualizations and graph-based reasoning, which provide insights into the decision-making process and enhance user trust. The system was trained and evaluated using benchmark datasets like UCSD Ped1 and Ped2, demonstrating high detection accuracy and robustness under occluded scenarios. Results showed that integrating occlusion-aware modules significantly improved the model’s performance compared to conventional methods. This research contributes to the field of intelligent surveillance by offering a reliable and interpretable solution that bridges the gap between deep learning advancements and practical deployment in real-world public safety applications.
DialyzeEase: A Web application designed to streamline the management of dialysis care for patients on hemodialysis and offer personalized educational modules
Chronic Kidney Disease (CKD) patients on hemodialysis face significant challenges in managing their care, including limited appointment availability, long waiting times, and difficulties in scheduling treatments. Additionally, research shows many patients lack crucial knowledge about how dietary and lifestyle modifications can slow disease progression. This project aims to develop DialyzeEase, a web application designed to address both challenges by streamlining appointment booking processes and providing personalized educational content to improve patients’ quality of life.
The research methodology followed a combined approach, beginning with a literature survey to validate the problem, background, and specific requirements. These findings were further evaluated through surveys with 41 respondents and interviews with two doctors from local hospitals. Stakeholder analysis using an onion model identified key actors in the dialysis care ecosystem, while system modeling established a robust design framework through functional, object, and dynamic models.
Findings revealed that most patients rely on manual booking at hospitals, with the majority reporting limited slot availability and long waiting times as major challenges. There is also a significant administrative burden on medical staff. Additionally, patients struggle with lifestyle management.
The developed prototype addresses these issues through an integrated appointment management system and a personalized knowledge hub that delivers CKD stage-specific recommendations.
ZenSearch: Revolutionizing E-Commerce Search through Advanced Multimodal Integration & Retrieval Techniques
The existing traditional e-commerce systems struggle to focus the user query to give a relevant product recommendation at the end of the retrieval stage where they mostly rely on unimodal approaches. This project explores this gap through developing an efficient multimodal retrieval system utilizing the ColPali architecture where the product images and captions are mapped effectively into a unified space to ultimately produce accurate and context-aware product recommendations.
Alz-InsightNet An Explainable Attention-Based Multimodal and Multimodel System for Early Alzheimer’s Detection
Alzheimer’s disease (AD) is a progressive and incurable neurological condition that presents major challenges for early-stage diagnosis. Conventional methods rely heavily on manual interpretation of MRI and PET scans, which can be subjective, time-consuming, and prone to error—often delaying timely intervention. This project addresses these limitations by developing Alz-InsightNet, an explainable, attention-based multimodal deep learning system for early detection of Alzheimer’s disease.
The system combines structural MRI and functional PET imaging data to improve diagnostic accuracy and support clinical decision-making. It employs two modified convolutional neural network (CNN) models—ResNet50 and DenseNet201—enhanced with Convolutional Block Attention Modules (CBAM) for MRI analysis, with their outputs integrated through an ensemble approach. For PET image classification, a CBAM-enhanced VGG-19 model is used. To foster clinical trust, the system incorporates multiple Explainable AI (XAI) techniques—Grad-CAM, Integrated Gradients, and LIME—that generate visual interpretations of the model’s predictions.
Alz-InsightNet demonstrated strong performance, achieving 98% accuracy with ResNet50, 95% with DenseNet201, 99.63% with the MRI ensemble, and 88% with the PET model. By fusing complementary imaging data and offering interpretable results, this system presents a practical and clinically relevant solution for enhancing the reliability and trustworthiness of early Alzheimer’s disease detection.