Projects
Semantic Segmentation with SWIN Transformers
• Implemented state-of-the-art using UNET, Transformers and transfer learning by fine-tuning model on 5k+ Cityscapes data
• Achieved significant improvement in mIOU score of 63%, utilizing SWIN attention residual mechanism with ML Perceptrons
Electricity Price Forecasting
• Forecasted daily and yearly prices using Timeseries analysis obtained by scrapping generation, consumption, weather data
• Feature engineered candidate variables using sliding window and applied Auto-Regression Differencing for reduced errors
• Achieved a low Mean APE of 9.69% for LSTM Model, outperforming the SOTA Kaggle model with 32% reduced (RMSE)
Question Answering model using BERT and derivatives
• Created a scalable QnA model by leveraging preprocessed Word2Vec, SIF embeddings on SQuAD v1.1 with 100K+ pairs
• Achieved high accuracy of 81% EM and 84.5% F1-Score by implementing Distil-BERT-BERT ensemble transformer model
Claim Prediction in Travel Insurances
• Developed an ensemble using boosted models to classify imbalanced claims data using feature selection and SMOTE analysis
• Utilized Flask framework to deploy the trained model as REST API, with prediction accuracy of 94.69% and F1-Score of 0.84
