Shweta Debjit
Sarkar

Two paired wildfire detection models exposing a critical ML lesson: a satellite classifier achieved 99.5% accuracy, but Grad-CAM revealed it learned land-use patterns, not fire damage a textbook shortcut learning failure. A second model on real fire/smoke imagery achieved genuine 100% accuracy, confirmed by interpretability analysis.

Binary classifier (healthy vs bleached coral) comparing VGG16 and EfficientNetB0 transfer learning on 923 images. Demonstrates the bias-variance tradeoff in small-dataset deep learning VGG16 overfit severely with 119M trainable parameters, while fine-tuned EfficientNetB0 achieved 79.5% accuracy with only a 2.1% train/val gap. Grad-CAM confirms ecologically valid feature attention.

7-class skin lesion classification on HAM10000, training two custom CNNs from scratch (CNN V1: 39.5% balanced accuracy; CNN V2: 42.1%) and benchmarking against ResNet50 transfer learning (83.8%). Demonstrates empirically why transfer learning is essential for medical imaging with limited data dermatofibroma recall was 0% for both scratch models vs 96% for ResNet50.

NLP pipeline detecting gender-coded, age-biased, and exclusionary language in 123,842 real LinkedIn job postings, grounded in the Gaucher et al. (2011) lexicon. Found 46% of postings lean masculine, with Tech and Venture Capital among the most biased industries a measurable gap between stated diversity commitments and actual hiring language.

Binary classifier (blank vs animal-present) for conservation camera trap images using fine-tuned ResNet18 on the Serengeti2 dataset. Data augmentation reduced the generalisation gap from 6.93% to 1.03%, achieving 85.11% test accuracy. Grad-CAM confirms the model attends to animal regions rather than background landscape essential for trustworthy deployment in wildlife monitoring.

Pilot classroom observation study (N=73, Years 7–11) across two Liverpool secondary schools examining whether students who actively interrogate AI output show higher conceptual understanding than those who passively copy it. Key finding: 32% of students claimed ownership of an answer they could not explain — a gap invisible to current assessment methods, consistent across both SEN and non-SEN populations.

Retrieval-Augmented Generation pipeline for question-answering over Parkinson's and Alzheimer's research literature. Built a semantic search and retrieval layer over a corpus of clinical abstracts, with retrieval evaluation metrics and query similarity analysis to assess answer grounding quality.

End-to-end metabolomics pipeline identifying urinary biomarkers of Type 2 Diabetes Mellitus from NMR spectroscopy data (MetaboLights MTBLS1, N=132). Combined ExWAS with Benjamini-Hochberg FDR correction and Random Forest classification (cross-validated AUC = 0.985) to surface 13 metabolites confirmed by both methods including hippurate and branched chain amino acids consistent with published T2DM literature.

Spatial epidemiological analysis linking environmental exposure data to cancer outcomes across English regions using ML and geospatial data linkage.

Binary classification of daily rainfall occurrence from multi-variable atmospheric observations humidity, pressure, wind direction, and temperature.

Clinical ML classification of malignant vs benign tumours using Logistic Regression, XGBoost, and Random Forest, with SHAP explainability for diagnostic feature identification.

Longitudinal cohort analysis on 100,000+ anonymised user records across 10+ countries. Identified 88% install-to-subscription drop-off and modelled country-level variation.