Integrative Deep Learning for Exoplanet Detection and Habitability Scoring
This project presents an AI-driven framework to accelerate the discovery and habitability assessment of exoplanets, bridging the gap between cutting-edge research and public participation. Leveraging light curve data from Kepler and TESS missions, the system employs a novel hybrid neural architecture that integrates Spiking Neural Networks (SNNs) with Long Short-Term Memory (LSTM) models for efficient and accurate detection of planetary transits, even in noisy data environments. The detection pipeline is followed by a habitability scoring module that computes key physical parameters such as planetary radius, orbital characteristics, and equilibrium temperature.
The platform is designed with accessibility at its core, featuring a no-code, user-friendly interface that allows users to search stellar data, visualize light curves, and confirm exoplanet candidates.
By combining state-of-the-art machine learning with usability and outreach, this project advances intelligent exoplanet detection and opens new pathways for collaborative scientific discovery, aligning closely with NASA’s mission to explore habitable worlds beyond Earth.
AMGAN: Attention-Driven Multi-Scale GAN with Complementary Learning Sub-Network for Generalized Non-Uniform Low-Light Image Enhancement
Low-light image enhancement (LLIE) is a critical component in computer vision pipelines, particularly in domains such as surveillance, autonomous driving, medical imaging, and photography. Images captured under poor illumination often suffer from low brightness, noise, and color distortions, which degrade the performance of downstream vision tasks. This project proposes AMGAN — an Attention-Driven Multi-Scale GAN integrated with a Complementary Learning Sub-Network (CLSN) — to address the complex challenges of LLIE, particularly under non-uniform lighting conditions.
The CLSN generates an inverse grey map that acts as an adaptive attention map, selectively enhancing underexposed regions while preserving well-lit areas. This attention-guided map is concatenated with the original low-light image to form an intermediate enhancement. The GAN’s generator, based on a U-Net architecture, extracts both spatial and frequency-domain features using Fast Fourier Transform (FFT), enabling fine-grained texture preservation and natural color consistency. A dual Markov discriminator ensures that both global and local image realism is maintained.
The model is trained and evaluated on benchmark datasets such as LSRW, LOLv1, LOLv2-real, and LOLv2-synthetic, achieving competitive performance in terms of PSNR, SSIM, and perceptual quality. Experimental results show that AMGAN balances enhancement quality with computational efficiency, making it a viable solution for real-world LLIE applications. Future work will explore architectural optimizations and potential domain-specific extensions.
Equity Portfolio Builder: Minimum Variance Portfolios for the Colombo Stock Exchange Stocks
Since the post-pandemic market rally, local and foreign investor participation in the Colombo Stock Exchange (CSE) has grown steadily. Despite this growth, the CSE remains undervalued compared to the regional peers, with a price-to-earnings (P/E) ratio of approximately 8x and a market capitalisation to GDP of 19%, signalling significant untapped potential for long-term capital allocation. One of the major causes for this is the lack of retail investor participation. Retail and novice investors face challenges in constructing diversified, risk-minimised portfolios due to low liquidity, the lack of awareness in the Sri Lankan capital markets, and the lack of tools and applications for portfolio creation and management. They often rely on ad-hoc advice or personal sentiment, resulting in suboptimal returns and heightened exposure to market fluctuations.
This project proposes the ‘Equity Portfolio Builder’ web application that empowers retail investors to build and manage risk-optimised equity portfolios across all CSE-listed companies. Grounded in Markowitz’s Modern Portfolio Theory, the application uses a quadratic optimisation engine to compute optimal weight allocations, minimising portfolio variance by analysing historical price data to generate covariance matrices. It features automatic portfolio rebalancing to adapt to changing market conditions and user goals. It is delivered through an intuitive, transparent interface that simplifies complex statistical processes to be accessible to average retail and novice investors. This tool democratises sophisticated portfolio construction, enabling retail and novice investors to make informed, data-driven decisions, bridging a critical gap in Sri Lanka’s capital markets by offering a cost-effective, self-directed solution to enhance investment management and encourage market participation.
Digital Mentor
Digital Mentor is an innovative AI-powered onboarding platform designed to significantly enhance employee integration within organizations. It addresses common onboarding challenges, including high employee turnover rates, inconsistent training quality, and prolonged periods before achieving optimal productivity. Leveraging advanced artificial intelligence technologies, Digital Mentor delivers personalized learning pathways tailored specifically to individual job roles, ensuring rapid knowledge acquisition and retention. Its sophisticated features include interactive micro-learning modules, AI-generated content summaries, and custom quizzes that reinforce learning and assess employee comprehension effectively. Additionally, the platform integrates a real-time AI chatbot trained on company-specific content, providing instant and reliable support to new hires, thus significantly reducing dependency on human mentors. Comprehensive administrative controls streamline content management, employee progress tracking, and insightful analytics, dramatically improving operational efficiency and decision-making capabilities. Future enhancements include advanced analytics dashboards, mobile application development, gamification elements, and seamless HR system integrations, positioning Digital Mentor as a leading-edge solution. Overall, Digital Mentor not only transforms onboarding processes but also empowers organizations by fostering rapid employee productivity, engagement, and satisfaction, creating a strategic advantage in employee management.
CLAIRO – Client Case Management System for Sri Lankan Law Chambers
This project is initiated to solve an issue faced by Sri Lankan legal sector, a digital solution that is designed to manage client and case information in one platform. The majority in this sector rely on manual processes, and this leads to disorganization, and inefficiency in reminding clients and track case progress. These types of obstacles reduce any attorney’s productivity and increase stress.
The aim of this project was to design, develop, and evaluate CLAIRO, a Client Case Management System that was to be access through a web browser and was specifically designed for Sri Lankan Law Chambers. It provides features such as: centralizing client and case information management, automatic WhatsApp reminding system, a comprehensive dashboard and multilingual support(Sinhala, Tamil, and English). The system design has compatibility with the actual workflows of Sri Lankan legal practices.
CLAIRO is initially built for Small to Medium Law Chambers to practice in the civil sector. In future, the application may expand to Large Chambers and Firms with civil, criminal, and commercial sectors ,and this application can lead to great legal tech solutions in Sri Lanka.
CricXpert: A Hybrid Approach Combining Facial and Spatio-Temporal Gait Analysis for Enhanced Player Recognition with LLM-Based Statistic Generation
In the dynamic setting of T20 International (T20i) cricket, recognizing fielding players during the final overs poses major challenges due to poor lighting, occlusions, motion blur, and non-frontal views. CricXpert addresses this by introducing a hybrid AI system that combines computer vision with natural language processing to ensure accurate player recognition and intuitive access to player statistics.
The system follows a multi-stage recognition pipeline: it begins with Optical Character Recognition (OCR) using EAST and Tesseract to detect jersey text. If OCR fails, it falls back to facial recognition, employing MTCNN for detection, FaceNet for embedding, and SVM for classification—achieving 95.83% accuracy. To improve generalization, a spatial model extracts ResNet50 features and classifies them via a stacking ensemble of SVM, KNN, and logistic regression, resulting in 98.14% accuracy. For movement-intensive clips, a GRU-based temporal gait model identifies players using engineered pose features such as joint angles and step length, reaching 95% accuracy.
Beyond vision, CricXpert integrates a natural language interface for stat retrieval. Using GPT-4o with LangChain and Pydantic, it converts user questions into SQL queries. OutputFixingParser and schema-aware prompts minimize hallucinations, achieving an 85–90% query success rate on 100 test questions—outperforming LLaMA 13B and Gemini.
The system has been validated using performance metrics and expert feedback from national-level coaches. CricXpert’s modular design allows scalability to other match formats and sports, marking a significant advancement in real-time sports analytics and AI-assisted decision support.
ClassiSolve: A Unified Adaptive System for Categorizing and Predicting Resolution Times for Poorly Described Customer Support Tickets
Poorly described support tickets are a major challenge in customer service systems, often leading
to misclassification, inaccurate prioritization, and inconsistent resolution time estimation. Most
existing systems handle ticket categorization and resolution time prediction as isolated functions,
often relying on rule-based logic or static estimation methods, with limited ability to adapt to
changing support patterns. This research proposes ClassiSolve, a unified adaptive system that
simultaneously categorizes support tickets and predicts resolution times.
The project was conducted using an Agile project management methodology, following an
incremental and iterative approach that incorporates Feature-Driven Development (FDD) for
modular design and Test-Driven Development (TDD) for reliability. The solution employs a
supervised learning methodology, utilizing a single XGBoost model trained on preprocessed
historical ticket data to perform both ticket categorization and resolution time prediction.
To optimize workflow, ClassiSolve also integrates an automated agent assignment mechanism that
recommends the most suitable handler for each ticket based on predicted category and staff skill
alignment. Evaluation metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error
(RMSE) confirmed the model’s accuracy and reliability. Real-world relevance was validated
through expert interviews and surveys among IT professionals, revealing a strong industry need
for adaptive systems.
ClassiSolve’s dynamic learning capability enhances support team efficiency and accuracy by
enabling continuous improvement. Observations indicate improved workload balance, faster
resolution, and more consistent service quality, while conclusions confirm the effectiveness of a
unified, data-driven model in addressing ticketing inefficiencies. The system offers a scalable and
practical solution for modernizing IT support operations.