Fr. Conceicao Rodrigues College of Engineering

Skin cancer is a serious public health challenge, and early diagnosis is crucial for effective treatment. We present a machine-learning pipeline to classify patients as “Cancer” vs “Non-Cancer” using clinical and demographic features. The study uses a real-world dataset ($\mathbf{n}=\mathbf{7 0 4}$) with 20 predictors (age, gender, lesion site, lifestyle factors, etc.) and an extremely imbalanced target (696 non-cancer vs 8 cancer cases), fortunately, due to less cancer patients observed in India. We perform thorough preprocessing (including handling missing values and categorical encoding), exploratory data analysis, and compare five classifiers: Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine (RBF kernel), and K-Nearest Neighbors. A stratified 5-fold cross-validation ensures reliable model comparison and overfitting control. The best model (Logistic Regression) achieved near-perfect cross-validated performance (ROC-AUC $\boldsymbol{\approx} \mathbf{1. 0}$) and very high test accuracy $(99.3 \%)$, with precision 66.7% and recall 100% on the rare cancer class. Performance metrics (accuracy, precision, recall, $\text{F1}$, ROC-AUC), confusion matrix, ROC curve, and feature importance (from the tree-based models) are reported. We include visualizations from EDA (class balance, age distribution, lesion site distribution, correlation heatmap) and evaluation (cross-validation bar chart, confusion matrix, ROC curve, and feature importances). The study highlights the predictive power of routine clinical features and provides insights for interpretable skin cancer risk assessment using SHAP and LIME.

Deep learning classifiers exhibit susceptibility towards iterative adversarial perturbations, often under high-fidelity attacks experiencing total categorical collapse. To address this, we introduce the Asymmetric Latent Purifier (ALP), a novel structural defence mechanism inspired by the stochastic information bottlenecks of the 2026 Unified Latents (UL) generative framework, Unlike Traditional deterministic autoencoders, ALP incorporates an adaptive, non-differentiable Gaussian noise layer within a 64-channel latent manifold to disrupt adversarial gradient flows. Empirically validated on CIFAR-10 dataset using an Apple M4 8-core GPU architecture. While the unprotected baseline experiences a total categorical collapse ( 0.00% accuracy) under a 7-step iterative PGD attack, our 20-sample adaptive ensemble approach achieves a robust accuracy of 32.06% (SD=1.94%)( averaged over 5 trials ) while ensuring a high-fidelity reconstruction of 25.68 dB. Operating a total system latency of 13.86ms, offers a promising path towards real-time flexibility for complex RGB varieties. Furthermore, with a single-sample inference latency of 1.25 ms, ALP represents a 100x to 1000x speedup over iterative diffusion-based purifiers, enabling real-time adversarial immunity in safety-critical systems.

The project investigates the possibility of exoplanets being habitable using a combination of data sources such as the Habitable Worlds Catalog (HWC) and the NASA Exoplanet Archive reporting on thousands of exoplanets that contains data such as radius, density, distance from star, spectroscopy etc. The project intends to determine the habitability of these planets with the mentioned parameters. The project is going to utilize four models, a Decision Tree Classifier that maps planets into types, rocky, gaseous, and so on; K means, which is utilized for the amount of radiation received from a star and space; a convolutional neural network and multi-layer perceptron for analysis of graphs of the spectrographs of the planet to help estimate atmospheric compositions; a random output regressor to analyze and calculate the Earth Similarity Index (ESI) of the planet. The four models are combined, each assessing a specific characteristic of the exoplanet to provide a habitability score, which evaluates whether a planet has the capacity to sustain life the way we know it. For additional interaction and visualization of the information, we have implemented a universe sandbox, i.e. a simulation to assess the implications of characteristics of these planets in comparison to earth, such as weight difference using the simulation of a freely falling body. Through this sandbox you can utilize your own characteristics and play with it to dissect the impacts. The sandbox provides a fun interactive way to instill interest in kids educating them about astronomy and physics. Investigation of exoplanets was done to assess the variations between both bodies. A few exoplanets exhibit some elements to be habitable even considering they are light years away, although indication of habitability is a crucial step to examining the universe and ourselves. For future work short-term, we plan to expand with more exoplanets into the dataset and continue improving our sandbox to be more interactive and accurate.

Buying intent detection is a central challenge that cuts across several artificial intelligence domains, including natural language processing, knowledge representation, and decision-making systems. Inferring a user’s purchase intent correctly is of critical importance in real-world applications such as e-commerce, customer relationship management, and personalized recommendation systems, where it directly impacts conversion rates, user engagement, and operational efficiency. Over the years, a wide range of approaches have been proposed that span deep-learning models, knowledge-graph based reasoning, reinforcement learning, and more recently agent-based and multi-agent systems. This review systematically analyzes the evolution and efficiency of these paradigms with an eye toward real-world applicability. Deep-learning models based on transformers are among the most powerful detection models, given large labeled datasets, but offer little in terms of interpretability or robustness under domain shift. Methods leveraging knowledge graphs augment reasoning and interpretability by modeling structured relationships but come at significant construction and maintenance costs. Reinforcement learning introduces the ability to learn adaptively in sequential and dynamic environments but is sensitive to reward design and sample efficiency constraints. AI-agent systems provide more comprehensive autonomy and facilitate multi-step coordination tasks while introducing novel challenges related to reliability, latency, and operational overhead. We provide a unified taxonomy that synthesizes these approaches and compare them on several dimensions, including detection accuracy, quality of the induced reasoning, robustness to domain changes, and efficiency in the deployment setting. Unlike previous surveys, this work puts an emphasis on deployment-realistic benchmarks such as latency, cost of inference, quality of the grounded output, and tool success rate, supported by reported benchmarks and ablation studies. We conclude by pointing out open challenges regarding scalability, interpretability, and integration of symbolic and data-driven methods, giving practical insights for both researchers and practitioners while developing intent-aware AI systems.

Demand uncertainty is still a major problem for supply chain managers, who a lot of times have to make a decision between having too much inventory and running out of stock. This paper proposes a useful Monte Carlo Simulation (MCS) that links demand variability straightway to the cost and profit involved in inventory decisions. The method is deliberately uncomplicated - it employs straightforward simulation steps that can be executed in a spreadsheet - however, it is still able to deliver significant insights for the management. A Triangular distribution with the parameters (70, 100, 130) was used to depict the demand for a month, indicating that even rough estimates could still be quite close to the actual number. The framework through 10,000 simulated scenarios evaluates holding, stockout and total costs for different stock levels. The U-shaped cost curve yielded by this analysis pinpoints 112 units as the best stock level, equivalent to 36.9 units of minimum expected total cost and a service level close to 95