Pratik Sakode

• Data Scientist at Sterlite Technologies Jan 2016 - Present

● Lead Prediction: Led the development of an XGBoost classification model to predict the probability of winning or losing high-risk deals which lead to a noteworthy 10% rise in deal conversions.

● Collaborated closely with sales and marketing teams to define key metrics for the model, ensuring alignment with business goals. ● Presented the model and its findings to key stakeholders, securing their buy-in and fostering support for its implementation. ● Rod cylinder matching: Developed a SVM based model to assess compatibility between rods and cylinders for fiber optic production with 85% accuracy. ● Deployed the model on the GCP cloud using the Flask web framework, making it readily accessible for plant personnel. ● This streamlined approach increased plant worker productivity. ● Fault detection : Developed a CNN based model to detect defects in core rodes. ● Deployed the model on Google Cloud Platform (GCP) using Flask for a scalable and accessible solution. The model used pytorch libraries for accurate predictions ● This system enabled early fault detection in core rods, leading to a 10% cost saving for the plant. ● Plant Dashboard : Created a user-friendly dashboard on Looker Studio for plant workers. ● This dashboard integrates data from BigQuery tables populated by scheduled queries and views. ● The dashboard provides plant workers with real-time or up-to-date monitoring of various machine parameters. ● Product Quality Optimization : Investigated relationships between parameters at different stages of optical fiber manufacturing. ● Analyzed the connection between these parameters and various quality indicators of the final fiber. ● The goal was to identify these relationships to ultimately improve fiber quality control methods during production. ● Few Shot learning : Developed a complete end-to-end deep learning pipeline using few-shot learning with PyTorch for characterizing distributed fiber optic sensor data. ● This pipeline facilitates the comparison of sensor data from various deployment locations, even with limited data available at each site.