▪︎ Co-invented, architected and implemented the patent-pending AI for the first of its kind multipurpose sound recognition system that is improving the quality of life across deaf and hard of hearing communities and quality of care across care homes and assisted living facilities in the UK, USA and Europe.▪︎ Implemented end-to-end ML pipelines that run real-time predictions on more than 4 TB of audio streams every day. The pipelines are deployed and managed on AWS Lambda as Docker containers through GitHub CI/CD.▪︎ Delivered a fast, fault-tolerant and scalable backend architecture that is thoroughly cost and resource optimised, producing ML inferences within 2 seconds of triggering events. The architecture aligns business strategy and vision with technical best practices of compute optimisation and careful resource allocation across AWS services.▪︎ Designed a novel architecture to continually improve the ML models through customer feedback. The models are dynamically re-trained based on binary (Yes / No) responses fed by customers through the Earzz mobile app, thereby achieving an average F1-score > 95% across all 25 sound recognition models.

▪︎ Led the development of NLP framework for classification of clinical documents and extraction of numeric measurements from unstructured texts in HL7 messages.▪︎ Built and deployed containerised ML applications that include data ingestion and processing, feature engineering, training, cross-validation and unit testing for standardisation of clinical codes and grouping of patient encounters.▪︎ Collaborated with multidisciplinary teams, clinical SMEs and external stakeholders to drive initiatives on data governance, security, and compliance across analytics solutions.

▪ Introduced the first computational models for Topological Analysis on streaming data. The framework is employed in diverse applications such as network anomaly detection and the study of viral evolution.▪ Reduced time and space requirements of persistent homology, a principal tool in Topological Data Analysis, by approximately 1,000 times with a small error bound guarantee.▪ Expert in data stream mining with in-depth knowledge of stream clustering algorithms.▪ Developed a clustering algorithm that combines random projections with locality-sensitive hashing while preserving privacy of distributed data sources. The algorithm, implemented on Apache Spark cluster in Standalone mode, is at least 11 times faster and 15% more accurate than its closest competitor.▪ Key contributor in two Federally funded research projects worth $1M.

▪︎ Integrated and parallelised multiple models through a framework called FOQUS that aids in optimisation and quantification of uncertainty. The developed methodology reduces the run time of simulations by 60% on a 4-core local node.▪︎ Introduced an approach for the development of algebraic surrogates that characterise the flow rates in positive displacement rotary pumps.▪︎ Built interactive R-Shiny dashboards and analytics for displaying the response of Bayesian models that characterise the qualities of products and the behaviours of consumers.