Stable Diffusion from Scratch

Complete implementation of Stable Diffusion including VAE encoder/decoder, CLIP text encoder, UNet with cross-attention, and classifier-free guidance. Generated 512×512 images from text prompts via DDPM denoising.

LLaMA 2 from Scratch

Built LLaMA 2 from the ground up with KV-Cache, rotary position embeddings, grouped-query attention, and top-p sampling. Implemented BPE tokenizer and ran zero-shot generation on custom prompts.

GPT from Scratch

Transformer-based language model with multi-head self-attention, feed-forward layers, and autoregressive generation. Trained with AdamW; achieved validation loss of ~1.89.

Transformer Implementation

Full Transformer architecture with token/positional embeddings, multi-head attention, and a complete training pipeline. Integrated TensorBoard logging and automated checkpointing.

Vehicle Insurance Prediction

End-to-end MLOps pipeline for binary classification of customer insurance interest. Integrated MongoDB Atlas for data storage, Docker containerisation, AWS S3/ECR for model registry, and automated CI/CD via GitHub Actions.

Water Potability Prediction

• Structured with Cookiecutter scaffolding; reproducibility via DVC pipelines.
• Trained five models: Decision Tree, Random Forest, SVM, XGBoost, and k-NN.
• Tracked experiments with MLflow + DagsHub; logged confusion matrices per run.
• CI via GitHub Actions; feature importance logging included.

Flight Delay Analysis & Prediction

• Analysed 178,747 flight records; engineered features including operational arrival index, delay rate, and cyclical time encoding.
• XGBoost model: 78% accuracy and 0.86 AUC-ROC on 35,000 validation records.
• 81% recall on actual delays; average delay duration predicted within 5.2 min MAE.
• SHAP analysis identified departure hour, weather, and carrier as the top three delay drivers.

Credit Default Prediction

• Engineered max delinquency, delinquency streak, and credit utilisation from 29,461 customer records.
• Preprocessing: median imputation, deduplication, 60/40 split, SMOTE for class balancing.
• Ensemble models: 73% accuracy and 0.76 AUC-ROC; 67% recall on true defaulters.
• Deployed on 5,017 test records with a tuned 0.465 threshold for high-risk intervention.

Stock Sentiment Analysis

• Python pipeline to ingest, clean, and convert financial news into sentiment features.
• Trained and tuned ML classifiers to predict market direction from news sentiment.
• Backtested with QuantStats on historical price data: 10% annual return, 1.25 Sharpe ratio.
• Scalable backtesting engine processing intraday OHLC data with execution slippage benchmarking.