End to End Football Leagues Analytics    [Live Dashboard]

• Designed a robust football data analytics pipeline on both AWS and GCP, encompassing 950+ global leagues, ensuring automated data collection, preprocessing, and storage.
• Deployed a cloud solution emphasizing serverless architecture, automated scheduling, minimal overhead cost, real-time data processing, and seamless integration with visualization and machine learning tools.
• Orchestrated data warehousing solutions using GCP’s BigQuery and AWS’s Redshift, enabling robust SQL querying, seamless data transfer, and integration with Looker, Tableau and machine learning platforms like Vertex AI and SageMaker.

Twitter Streaming Sentiment Analysis    [Report]

• Built streaming Pipeline using Twitter API, NiFi, Kafka, and Spark Streaming using AWS EC2 Instance.
• Trained Word2Vec model and Decision Tree models for sentiment prediction. Achieved 78% accuracy, and 0.69 AUC-ROC.
• Built a dashboard for visualizing the sentiment on particular topic, and adhoc-queries on streaming data for 5 mins window.

SQL Queries performance evaluation    [Report]

• Extracted the IMDB dataset from server, preprocessed, and created the ER diagram, DDL and DML statements.
• Created different analytical queries, and evaluated the performance of lookup queries with and without indexing.
• By indexing for lookup queries reduce the queries time by about 85%.

Bank Telemarketing Term Deposit Subscription Analysis and Prediction    [Report]

• Performed the Data Cleaning, Exploratory data analysis, ANOVA, and chi-square test to check feature independence.
• Trained the logistic regression and decision tree classifier to predict the subscription probability. Achieved 0.60 AUC-ROC.

• Performed data preprocessing, exploratory data analysis, feature engineering, and treating the unbalanced data.
• Trained classifier a logistic regression and random forest classifier for predicting the annual income class. Achieved 0.66 F-1 Score and 0.88 ROC-AUC score.

• Implemented a deep implicit movie recommendation system for IMDB dataset to predict the rating of user for particular movie
• Optimized the model using triplet loss, by measuring similarity between user and item embeddings. Achieved 92% accuracy

• Developed a object detection model to identify the 80 different objects with 5 anchor boxes by using YOLO algorithm.
• Performed transfer learning using pre-trained YOLOv2 model and increased the accuracy of model by applying Non-max supression.

Remaining Useful Prediction for the Aircraft Gas Turbine Engine    [Report]

• Performed the exploratory data analysis, feature engineering, and build the CNN architecture for time-series analysis
• Trained the CNN model and fine-tune the model to overcome the overfitting issue. Achieved 21.5 MAE and 24.3 RMSE.

• Implemented semantic image segmentation on the CARLA self-driving car dataset.
• Used Encoder block and Decoder block along with skip-connection at each level to improve the accuracy of masking prediction.
• Applied sparse categorical cross-entropy loss to train the model for pixelwise prediction

• Trained Convolutional and Artificial Neural Network based classifier to predict the handwritten 28x28 pixel image.
• Achieved accuracy of 98% in CNN and 97% on ANN. Evaluated F-1 Score, Recall, Precision and confusion matrix.

• Developed an attention-based model for Neural Machine Translation (NMT) specifically designed to translate human-readable dates into machine-readable dates.
• Model incorporates pre-attention Bi-LSTM and post-attention Long Short-Term Memory layer to enhance translation accuracy.

• Developed a Transformer-based model to process resumes and extract information such as name, skills, designation etc.
• Performed transfer learning using the DistilBERT fast tokenizer and a pre-trained transformer model for parsing resumes.

• Implemented Viterbi algorithm for parts of speech tagging for each word in a sentence using Penn Treebank datasets.
• Achieved 92.3% accuracy on test set using Viterbi algorithm and 85.18% accuracy for assigning most frequent tag.

• Applied Naïve Bayes algorithm for identification of correct sense of particular word in given context.
• Created a bag of words for each sense for a given words in a training set.
• Implemented 5 fold cross validation and achieved 90.18% accuracy on ‘bass’ dataset.

• Developed the text summarization by using the concepts TF-IDF and Centroid on CNN News Dataset.
• Centroid based approach is providing better central idea for given text and achieved the average 0.42 ROUGE-1