UTSAV DATTA

Achieved 85% accuracy in ML based spend classification for a consumer goods company’s procurement workflow to quantify and improve their savings opportunities. Achieved over 40% reduction in redundant/duplicate data using Locality Sensitive Hashing and Jaccard distance techniques among others to reduce training data labelling effort. Designed and productionized a supplier news analytics module that included news categorization, sentiment analysis, topic modelling and entity recognition using an ensemble of Deep Learning techniques for a proprietary procurement management product. Designed / developed / productionized a proprietary product which extracts & formalizes unstructured information into knowledge representation using Python and Deep Learning models. Key role in improving product efficiency by transforming concepts from Deep Learning research papers into usable features.

Responsible for designing & development of an end-to-end machine learning model to predict outcome of new contract proposals for a US based print and digital document seller with baseline accuracy of 80%. Integral role in improving baseline accuracy of the above-mentioned model by 4% using Box Cox transformation, Gradient Boosting algorithm and Voting classifiers from Scikit-Learn. Engaged in creation of an end-to-end natural language processing system to identify and extract relevant portions from a large corpora of customer grievance emails for a Switzerland based insurance provider. Successfully achieved 70% F1-score (5% more than client expectation) using spaCy, NLTK, fastText, LSTM and Python regular expressions. Accomplished 50% faster comprehension of the above-mentioned emails thereby improving the overall turnaround time of the system. Instrumental in implementing a continuous model improvement pipeline in production to track system efficiency; gathered training data over time using qualitative & quantitative methods. Achieved optimum balance between performance, scalability and cost of production deployments of training / inferencing pipelines using both closed source (Azure Machine Learning compute clusters, Azure Kubernetes services) and open source (Docker, fastAPI) technologies. Established trackable and sustainable pipelines of above-mentioned ML models utilising Azure ML, Databricks and ML flow technologies.

Successfully created a business intelligence semantic models using SQL Server Analysis Services (SSAS) Tabular to facilitate advanced analytics on retail data. Involved in migration of 450+ reports to Power BI from legacy systems. Served as a Team leader (comprising of 5 team members), onsite coordinator and SPOC for US and APAC clients.

Developed Oracle PL/SQL packages procedures for ETL of telecom inventory data from legacy source systems to a modern data warehouse to support better service assurance.