Publications & Intellectual Property

First test of whether JSON-schema output distorts LLM confidence. A paired-bootstrap evaluation (ECE, Brier, AUROC) over 900 paired observations on Claude and Gemini found no calibration cost in 9 of 10 comparisons, but token cost up to 8× on one vendor. The evaluation also caught an API sampling restriction and a scoring artefact that would have distorted results. Whole study: US$3.24.

A harness scoring 6 explainers (SHAP, LIME, Integrated Gradients, DiCE, GNNExplainer, LLM rationales) on 4 metrics across 5 tabular, graph and text domains. Gradient attribution recovered the ground-truth graph motif perfectly where GNNExplainer averaged 0.556, and a widely used XAI toolkit failed on all 5 domains.

A resource-aware pipeline fine-tuning XLM-R for Hindi and Marathi. Quality-gated LLM augmentation raised Marathi F1 by 21.2 points with only 100 labelled sentences, for US$2.28, while Hindi gained at most 1.4 points and simple rule-based augmentation did best. Augmentation value depends on how scarce a language's resources are. Includes a blinded native-speaker evaluation.

A 2⁴ factorial multi-agent simulation of feed ranking, LLM sycophancy, coordinated accounts and a safety guardrail, studying how these effects compound. A calibrated-mock LLM method made full simulation runs nearly free.

Systematic review of why over-exercising athletes are at heightened SCD risk, how COVID-19 compounds cardiac vulnerability, and how IoT wearables can enable real-time monitoring and early intervention.

Applies ML regression and time-series forecasting to historical grid data to quantify state-level energy demand-supply gaps across India, supporting data-driven infrastructure planning.

Proposes a generative AI pipeline that synthesises stylised sticker images directly from free-text prompts, eliminating manual design effort for personalised visual content creation.

Compares anomaly-detection and classification models on SCADA sensor streams to predict turbine faults before failure, cutting unplanned downtime and preventive maintenance costs.

Surveys CNN and transformer architectures applied to MRI and fMRI scans for neurological disorder classification, benchmarking performance trade-offs across dataset scale and modality.

Integrates soil-moisture sensors, an ML-driven irrigation scheduler, and mechanical water-delivery hardware into a closed-loop IoT system for autonomous, optimised plant growth.

Presents a unified framework that chains automated EDA — distribution analysis, correlation, outlier detection — with AutoML model selection, shrinking time-to-insight from hours to minutes.

Combines Nash-equilibrium game theory with ARIMA-based forecasting to model how competing industrial stakeholders respond to ozone regulation, proposing incentive mechanisms that reduce emissions without coordination failure.

Developed a real-time Sudden Cardiac Arrest (SCA) detection system combining fog computing, IoT wearables, and a Multilayer Perceptron (MLP) deep learning model, achieving 99.52% accuracy — outperforming prior benchmarks — with an integrated alert system to prompt timely medical intervention.

Benchmarked three neural network architectures (basic, deeper, and wider) on the MIT-BIH Arrhythmia Database for QRS complex analysis and cardiac abnormality prediction, with the wider network achieving the best performance at 94.84% accuracy and an F1 score of 0.9457, demonstrating superior detection of subtle ECG patterns for early clinical intervention.

Registered design for a smart drip-irrigation device that reads soil-moisture sensors and autonomously triggers water delivery, reducing manual effort and water waste for home and small-scale horticulture.

Registered design for a wearable cardiac monitoring device that continuously tracks physiological signals and triggers an emergency alert when sudden cardiac death indicators are detected.