Worked as a part-time Full Stack Engineer on contract
Contributed to Case Study Intake (CSI), an internal Novartis reporting tool providing operational metrics on clinical trial start-up, conduct and closure phases
Continued working on the AEBrain project, the Novartis internal potential adverse events (AEs) detection system
Key contributions:
Reduced AWS costs by ~50% by migrating CSI data processing pipelines from AWS Sagemaker notebooks to scheduled Fargate containers
Improved CSI application management experience via a Python Dash based dashboard, allowing for easier ad-hoc pipeline executions and configuration changes
Refactored the AEBrain codebase, focusing on a modular architecture for faster and more dynamic onboarding of new use cases while maintaining backward compatibility
Migrated legacy model serving logic to TorchServe serving as part of the MLOps modernization, allowing for a fully automated model deployment pipeline
Worked as an Associate - Full Stack Engineer
Contributed to the AEBrain project, a GxP-compliant pharmacovigilance detection system that automates the monitoring of potential adverse events (AEs) mentioned in text aggregated from different sources using text classification machine learning models
Worked on initiatives like the Model Catalog, an internal web portal for Novartis engineers to search and reuse existing pre-trained ML/NLP models already developed by other Novartis in-house projects that are shared on this platform
Key contributions:
Scaled monitoring scope of AEBrain from just English-language text to supporting six languages from eight countries aggregated from 4 different systems within two years
Implemented a rule-based Expert system to complement machine learning predictions, reducing falsely flagged text by AEBrain down to 20% while maintaining 99% recall
Contributed to the migration effort of AEBrain from the existing server-based to full serverless on Amazon Web Services (AWS), with an estimated 35% reduction in operational costs
Modernized the AEBrain Jenkins CI/CD pipeline from legacy freestyle-based pipelines to Jenkinsfile-based pipelines, achieving higher reliability and about 40% reduction in build time
Developed a near real-time monitoring and alert dashboard for AEBrain using Python Dash to facilitate the activities of the service operation team
Developed the Model catalog web portal APIs using Python Falcon, designed the backend DynamoDB database schemas and implemented search functionality using Elasticsearch
Helped design the AWS cloud infrastructure using Serverless Framework for Model Catalog