SARANG DEB SAHA

Collaborated with Data Engineering and Data Science teams to design scalable dashboards and automate data pipelines.

• Collaborated with Data Engineering and Data Science teams to design scalable Apache Superset dashboards and automate data pipelines using Python, SQL, and Airbyte, optimizing task scheduling with DAGs.

Developed and fine-tuned deep learning models for real-time sports ball tracking.

• Developed and fine-tuned deep learning models (TrackNet) for real-time squash ball tracking using Python, TensorFlow, PyTorch, and OpenCV.
• Generated comprehensive datasets and processed match footage with FFmpeg, yielding key insights into player behavior and match dynamics.
• Deployed models in Docker environments and collaborated with cross-functional teams to integrate tracking data into broadcasting systems, significantly enhancing viewer engagement.

Leading the end-to-end development of a flagship production tool for financial technology.

• Led end-to-end development of 'FinSpector', a flagship production tool for bank statement parsing, transaction categorization, and mule detection, positioning it among a select few platforms in India.
• Engineered a robust PDF and OCR pipeline leveraging regex and NLP to achieve 97%+ accuracy in transaction extraction and classification, deploying a rule-based fraud engine to identify suspicious patterns.
• Enabled clients to perform real-time creditworthiness assessment and fraud detection directly from uploaded statements, reducing manual underwriting time by 60%.
• Positioned the tool for enterprise rollout with SaaS pricing, forecasting significant revenue opportunities through B2B lending partnerships.

Conducted a comparative study on road safety datasets and implemented computer vision models for traffic violation detection.

• Conducted a comparative study of road safety datasets and blackspot definitions across India, UK, US, and France, identifying critical feature gaps and inconsistencies in Indian data.
• Implemented YOLO-based computer vision models with centroid tracking to detect traffic violations including triple riding, helmet-less riding, and wrong-way driving.
• Developed a web scraping tool using BeautifulSoup4 to extract and structure FIR data from Karnataka (post-2016) via bounding box-based parsing.
• Performed spatio-temporal analysis of violations across 50 Bengaluru traffic junctions, integrating IUDX, BTP, and meteorological datasets to identify violation trends and build predictive models.

Contributed to the development of a recommendation engine and improved real-time player tracking solutions.

• Developed a recommendation engine to automate storyline generation for cricket commentators.
• Tested and benchmarked the company's real-time player tracker solution (QT and QStat), improving efficiency by 20%.
• Contributed to the development of new YOLO-based object detection models for cricket ball tracking, leveraging highlight videos for real-time detection and tracking in QT (Quidich Tracker).

Exploratory data analysis and predictive modeling in football analytics, focusing on Premier League teams and the transfer market.

Developed a real-time traffic density estimation system using YOLOv8 for urban traffic management.

Explored the IPL auction market using statistical analysis and machine learning models to uncover insights into player valuation and predict auction prices.

Paper accepted for presentation at ICRTCIS 2024 conference and published in the Springer Book Series 'Algorithms of Intelligent Systems'.

Paper on football analytics currently undergoing peer-review for publication in a Q2 journal.

Awarded prestigious fellowship as the only shortlisted team from Jaipur.

Achieved a top 6 spot out of 20,000+ students in Rajasthan zone finals and 1st in Cluster wildcard round.

Research on detecting anomalous events with criminal intent using deep learning, specifically Convolutional Neural Networks.

Exploration of the IPL auction market using statistical analysis and machine learning models to uncover insights into player valuation and predict auction prices.