Data-Driven Optimization of Call Center Operations for United Airlines, NLP [🏆]

As part of this project, I analyzed United's call center data to identify factors affecting prolonged call durations, including agent performance, call types, and customer sentiments. Using insights from this analysis, I proposed solutions to optimize operations, such as upskilling agents, redistributing call loads, and designing a more efficient Interactive Voice Response (IVR) system. These changes aimed to address frequent customer issues like refunds and baggage inquiries effectively.

To support these optimizations, I employed machine learning models, including a fine-tuned BERT-based classifier and classical algorithms like AdaBoost, for predicting call reasons. Despite challenges such as dataset imbalance, these models enabled better issue categorization and resolution strategies. Additionally, I developed methods to prioritize urgent calls through automated keyword-based classification, improving queue management and customer satisfaction. The project concluded with actionable suggestions to refine call center processes and maintain continuous improvement through customer feedback mechanisms.

Attribute-Value Prediction from E-Commerce Product Descriptions, NLP [🏆]

In this project, I worked on addressing a critical challenge in e-commerce: predicting attribute-value pairs from unstructured product descriptions. This task involved developing models to classify product details, such as brand and hierarchical categories (L0–L4), to enable efficient search, recommendation, and customer query resolution. I explored multiple approaches:

Generative Models with T5 (small/base): Leveraged Transformer-based encoder-decoder architectures to predict brand and category values. Post-processing techniques were employed to address generative hallucinations, improving the F1 score for the brand attribute from 0.3686 to 0.4601.

Trigram-Based Similarity Model: Applied a custom text similarity function to categorize hierarchical attributes by leveraging weighted trigram comparisons. This efficient k-nearest neighbors-like approach required no retraining when new data was added.

FastText Embedding Classifier: Designed a shallow neural network using character n-grams for semantic understanding. Though experimental challenges impacted replicability, the approach was efficient and well-suited for handling spelling variations in product descriptions.

Through these methods, the project tackled issues of noisy and imbalanced data, achieving structured results for downstream applications in search and recommendation. Detailed computational setups, hyperparameters, and reproducible architectures were documented, supporting ongoing innovation in metadata extraction for e-commerce platforms

Knowledge Graph Construction of Indian Legal Documents, NLP

Enhancing the construction of high-quality triplets for knowledge graph creation (KGC) in Legal domain.

Supervisors: Dr. Vasudha Bhatnagar (Senior Professor, DUCS) and Dr. Vikas (Assistant Professor, DUCS).

Deepfake Detection and Temporal Localization, Computer Vision

Established server infrastructure from scratch and conducted a comprehensive literature review and analysis of existing SOTA architectures, including UMMAFormer, BA-TFD, BA-TFD+.

Successfully reproduced results of 3 research papers detecting deepfakes, using NVIDIA A100 80GB GPU.

Contributed to the UMMAFormer’s GitHub repository, by updating required missing packages & correcting the file structure flow chart provided by the author in the README file.

Supervisors: Dr. Vasudha Bhatnagar (Senior Professor, DUCS) and Dr. Bharti (Assistant Professor, DUCS).

FoG Detection in Parkinson's Disease, AI in Healthcare

Detected which specific events trigger freezing of gait (FOG) to occur in patients having Parkinson's disease.

During a FOG episode, a patient’s feet are “glued” to the ground, preventing them from moving forward despite their attempts. FOG has a profound negative impact on health-related quality of life — people who suffer from FOG are often depressed, have an increased risk of falling, are likelier to be confined to wheelchair use, and have restricted independence.

Tested several machine learning models trained on data collected from wearable 3D lower back sensors.

Sign Language Detection System, Computer Vision

Developed a sign language detection system using transfer learning with TensorFlow’s SSD mobilenet v2 pre-trained on the Microsoft COCO dataset for object detection.

Created a custom dataset with ten annotated images per class to fine-tune the model, integrating OpenCV for image capture.

The model exhibited real-time classification accuracy on a live video feed, validated through successful peer evaluations, highlighting its adaptability to diverse scenarios.

Completed under the guidance of Dr. Punam Bedi (Senior Professor, DUCS).

Sentiment Analysis using Knowledge Graph, NLP [🏆]

Made a WebApp using Flask which can scrape the Tweets on various parameters (such as keyword, since, till, count, etc.), cleans them using RegEx.

Analyses and Classifies their Sentiments using TextBlob library, visualizes the results using matplotlib.

Stores this classified data into neo4j database to obtain Knowledge Graph for further Querying and Analysis

House Price Prediction

Started with this project as my stepping stone into the world of Data Science. Motivated by the 5-day long Deep learning workshop attended recently back then.