AeroSynth: A Physics-Informed Multi-Resolution Graph Framework for Aircraft Fuel Prediction and Route Optimization
“The aviation industry faces continuous challenges in balancing operational efficiency with accurate fuel prediction and route planning, largely due to a critical lack of publicly accessible, tail-specific performance models. Operators are often forced to rely on generic manufacturer specifications that ignore individual engine degradation over time, while traditional flight planning software depends on rigid airways that struggle to dynamically bypass volatile weather. Furthermore, alternative deep learning prediction methods are plagued by opacity and a lack of physical constraints, preventing the critical trust required for flight-deck certification. To overcome these barriers, this project introduces AeroSynth, a physicsinformed machine learning framework integrated into a realtime route optimization system that fuses data-driven graph intelligence with core aerodynamic constraints. The implemented solution features a novel PhysicsInformed MultiResolution Sensor Graph KolmogorovArnold Network (PIMRSGKAN) trained on extensive Quick Access Recorder (QAR) data. Utilizing a "residual-corrector" loop, the model maps 25 distinct sensor streams into an adaptive graph network, blending physical principles with wavelets and learnable spline activation functions to capture transient sensor dependencies across all flight phases. This predictive engine is integrated into an interactive decisionsupport web application that continuously streams live operational parameters including dynamic weather feeds, winds, active FAA NOTAM alerts, and restricted airspace boundaries to evaluate and contrast optimized direct, layover, and multistop flight tracks. Rigorous empirical testing on 410 held-out, flight-level test logs demonstrates that the PI- MRSG-KAN framework achieves exceptional predictive fidelity, securing a coefficient of determination (R 2 ) of 0.9814 and a Symmetric Mean Absolute Percentage Error (SMAPE) of 10.63%. Crucially, ablation testing reveals that anchoring the system with a physics-causal graph topology drops the Mean Absolute Error (MAE) by 44% compared to standard non- graph flat feature vector benchmarks. Furthermore, qualitative evaluation by domain experts confirms that the application effectively delivers clear, context-specific explanations of its route updates, successfully reducing pilot planning workloads. This integration validates that physicsinformed learning can be operationalized into a practical, interpretable, and open- source tool, bridging the gap between highfidelity prediction and real-world usability for aviation stakeholders.”
DYNAMIC OPEN WORLD OBJECT DETECTION: INTERGRATING INCREMENTAL FEW-SHOT LEARNING
“Traditional object detection models operate under a closed-world assumption, meaning they can only recognize the fixed set of classes they were trained on and typically misclassify any novel, unseen object as background noise. This creates a significant gap for real-world deployment, where systems inevitably encounter objects outside their training distribution. Open-World Object Detection (OWOD) has emerged as a research area to address this by enabling models to both identify known objects and flag unexpected ones, while also incrementally learning to recognize these new objects over time without forgetting previously learned classes. "Dynamic Open-World Object Detection (D-OWOD)," was scoped to advance OWOD toward practical, real-world application by designing a system that could not only detect unknown objects but also automatically extract their distinguishing features, group similar unknown objects into dynamically formed clusters, and incrementally incorporate these new clusters into the model using few-shot learning — all while avoiding catastrophic forgetting of existing knowledge. Given the project's timeline and resource constraints, this scope was ultimately refined.”
Bias Lens
“Bias Lens is a research project aimed at developing an advanced system to detect and explain various biases in written text. The goal of Bias Lens is to enhance fairness and transparency in language processing applications by integrating Explainable AI (XAI) methods. This project will identify multiple types of biases—such as gender, racial, and ideological biases—and provide clear, understandable explanations for these detections. This ensures that both technical and non-technical users can easily interpret and address biases in their work.”
MultiCardioX - Explainable Ensemble Deep Learning Framework for Multi-Label Classification of Cardiovascular Disease associated Comorbidities
“Artificial Intelligence (AI) is a growing trend in Cardiology to help in diagnosis. Many of the existing models, however, are only applicable to categorize the patients under a binary classification method and do not show the medical complexity of patients who have Cardiovascular diseases (CVD) and comorbid conditions such as Hypertension, Diabetes, Chronic Kidney Disease, etc. Furthermore, the black-box nature of the complex deep learning models might be a hindrance to clinical practice, as convincing diagnostic justifications are what medical professionals need in order to make critical decisions. To help mitigate them, the proposed research presents MultiCardioX, an interpretable ensemble deep learning model that addresses the multi-label problem of cardiovascular disease and its related comorbidities. It is founded upon an intricate stacking ensemble framework where XGBoost, TabNet, and MLP (Multi-Layer Perceptron) are called upon, along with a soft MLSMOTE balancing approach, to solve the issue of label imbalance. The system also brings in transparency by employing SHapley Additive exPlanations (SHAP) to offer fine-grained featurespecific explanations to its multi-label predictions. This framework was evaluated and got a Macro-Average AUC-ROC of 0.8379. The findings showed that the ensemble approach improved the detection of co-occurring conditions as compared to a single architecture. This study is also a significant advancement in the making of strong clinical decision support systems in complex cardiovascular care.”
AEGIS: A Compliance-Aware Multi-Agent System for Governed AI-Assisted software Requirements Engineering
“AEGIS is a multi-agent framework for requirements engineering. It extracts requirements from analyst text, scores vagueness and asks clarifying questions, then checks each against GDPR and EU AI Act rules via retrieval-augmented generation, returning explainable evidence, fixes, and an audit trail.”
ChainGuard
“Blockchain technology is known for its transparency, immutability, and decentralized trust. Still, there are two main challenges limit its adoption. First, there is a trade-off between privacy and transparency. Second, storing large or sensitive data directly on the blockchain is difficult. Public blockchains can expose data and compromise confidentiality, while storing data on- chain can cause problems with scalability and cost. This study suggests ChainGuard, a hybrid blockchain-based system that combines private and public blockchain networks, as a solution to these issues. While only cryptographic hash values are recorded on a public blockchain to guarantee transparency, immutability, and verifiability, sensitive data is safely stored in a private blockchain. To enable safe data retrieval and restricted access, the system also integrates private key-based access control and UUID-based identification. In order to ensure seamless interaction between system components, the solution is implemented using Nodejs for backend services and smart contracts deployed via Remix. The results demonstrate that the proposed system successfully preserves data privacy while maintaining transparency and integrity. The system provides secure access control, reliable data verification, and resistance to unauthorized access and data tampering. Overall, the proposed approach offers a practical and scalable solution for real-world applications requiring secure and privacy-preserving data management.”
Sherlock
“The increasing reliance on software in various applications has made the problem of software vulnerability detection more critical. Software vulnerabilities can lead to security breaches, data theft, and other negative outcomes. Traditional software vulnerability detection techniques, such as static and dynamic analysis, have been shown to be ineffective at detecting multiple vulnerabilities. To address this issue, this study employed a deep learning approach, specifically Convolutional Neural Networks (CNN), to solve the software vulnerability detection problem. A 5-split cross-validation approach was used to train and evaluate the CNN model, which takes tokenized source code as input. The findings indicated that Sherlock successfully detected multiple vulnerabilities at the function level, and its performance was particularly strong for CWE-199, CWE-120, and CWE-Other, with an overall high accuracy rate and significant true positive and true negative values. However, the performance was less reliable for some vulnerabilities due to the lack of a standardized dataset which will be a future research direction. The results suggest that compared to current techniques, the proposed deep learning approach has the potential to substantially enhance the accuracy of software vulnerability detection.”