Bikram Saha

babyGPT is a transformer-based language model built from scratch using PyTorch, featuring 275 million parameters with 16 layers and 12 attention heads. Trained on the Wikitext-8 dataset, it includes a custom tokenizer and employs the AdamW optimizer for improved convergence. The model incorporates core components like multi-head self-attention, feedforward layers, layer normalization, and positional encoding, all implemented manually. After over 15,000 iterations, it showed steady loss reduction, indicating successful learning. Teengpt provided deep insight into large language model design, training, and optimization, and reflects my passion for AI and hands-on expertise in deep learning architecture.

IRCTC - Backend Service Java-based backend service designed to automate IRCTC train ticket booking workflows, built using core Java and a lightweight LocalDB setup. The system implements modules for ticket booking, cancellation, seat availability tracking, and session management without relying on external database systems, enabling faster local data access and reduced overhead. It features custom data models and logic to manage real-time seat allocations and cancellations efficiently. The application follows a modular and object-oriented design, optimizing performance for single-machine deployments and enabling future scalability toward networked or RESTful implementations.

Walmart Sales Data Analysis  end-to-end pipeline is developed for Walmart sales forecasting using real-world time-series data. The dataset is preprocessed to handle missing values, convert date features, and engineer variables such as holiday flags and rolling averages. Exploratory Data Analysis (EDA) is performed to uncover store-wise and department-level sales trends, seasonality, and anomalies. Machine learning models including Random Forest Regressor and XGBoost are trained on historical data with hyperparameter tuning and cross-validation. Feature importance analysis highlights the impact of temporal and event-based variables. The final models are evaluated using RMSE and MAPE, delivering robust short-term sales predictions for operational optimization.

Paddy Disease Classification  A Vision Transformer (ViT-Large-Patch16-224) model is fine-tuned to classify paddy leaf diseases across 10 distinct categories using high-resolution agricultural imagery. The workflow includes rigorous data preprocessing: resizing to 224x224 pixels, random cropping, affine transformations, and normalization based on ImageNet stats. Training is conducted in PyTorch with the AdamW optimizer and cross-entropy loss. Careful tuning and augmentation strategies help the model generalize well, achieving 98.15% accuracy on the Kaggle competition test set. Evaluation includes per-class metrics and confusion matrix analysis, demonstrating the viability of transformer-based architectures in real-world plant disease classification and precision agriculture.