Piyush
Maji

Addressed the challenge of detecting fraudulent Bitcoin transactions, where traditional ML models ignore transaction relationships, by developing a Graph Neural Network pipeline that leverages graph connectivity for node classification. Engineered a graph learning workflow on the Elliptic Bitcoin Dataset (203K+ transactions, 234K+ edges) using GraphSAGE and Graph Attention Networks (GAT), with preprocessing, class balancing, and explainability through SHAP and GNNExplainer.

Built a production-ready inference pipeline by exposing fraud predictions through a FastAPI REST API and containerizing the application with Docker for scalable deployment.

Solved the problem of retail demand forecasting by building a scalable SKU-level sales prediction pipeline using LightGBM, enabling accurate inventory planning across multiple stores. Designed advanced time-series features including lag, rolling-window, calendar, promotion, and holiday variables, and benchmarked the model against SARIMA and Prophet baselines using time-aware cross-validation.

Implemented hierarchical forecast reconciliation to maintain consistency across SKU, store, and regional forecasts — improving the usability of predictions for supply chain and inventory decision-making.

Implemented Multinomial Naive Bayes (MLE + Laplace smoothing) and Logistic Regression (gradient descent) entirely from scratch in NumPy on the UCI SMS Spam Collection (5,574 messages). Built both TF-IDF and Bag-of-Words feature pipelines without sklearn, with vocabulary constructed strictly on the train split to prevent data leakage on an 86.6% class-imbalanced dataset.

Hand-coded binary cross-entropy loss and gradient descent with real-time loss convergence plots. Evaluated both models via precision, recall, F1, and ROC/AUC (computed using the trapezoidal rule). Achieved 94.9% F1-score and 0.982 AUC on the UCI SMS Spam Collection, demonstrating strong performance through threshold optimization and comparative model analysis.

Implemented and compared three compartmental epidemic models — SIR, SEIR, and the SUTRA model (which explicitly tracks undetected/asymptomatic carriers via a 'Unreported' compartment) — on US national COVID-19 data from Feb 2020 to Mar 2021. Used NNLS regression for automatic per-phase parameter estimation of contact rate β̃ and reach ρ̃, with a product-R² fallback for collinear phases, achieving MAPE as low as ~4% on individual phases.

Ensembled all three models via linear regression for improved overall fit. ODE systems solved with scipy's RK45 integrator. Visualised phase transitions, reproduction numbers, and ensemble predictions across the full pandemic timeline.

Built a two-script Python pipeline to extract laser wavelength from camera images of diffraction gratings. Stage 1 preprocesses raw frame stacks into 1D intensity profiles via row-wise averaging and background subtraction. Stage 2 fits the full Fraunhofer grating equation to the resolved diffraction peaks via non-linear least squares (scipy.optimize.curve_fit), yielding λ = 524 ± 6 nm — consistent with the known green laser standard.

Simulated 2D stochastic particle trajectories under Brownian dynamics using the Langevin equation with Gaussian noise. Computed the Mean Squared Displacement (MSD) ensemble average across thousands of trajectories and fitted MSD ~ t^α via log-log regression to classify diffusive regimes: sub-diffusion (α < 1), normal diffusion (α = 1), and super-diffusion (α > 1). Validated against Einstein's analytical MSD prediction (MSD = 2dDt), confirming the diffusion coefficient D within numerical precision.