Akshay Kumar

Models migration to Vertex-AI from Datarobot AutoML. Migrated models from Datarobot to Vertex-AI ensuring scalability, retaining logging & monitoring via Vertex-AI (both v1 & v2 version of monitoring). Optimized Training infrastructure with KubeRay. Implemented KubeRay on GKE to parallelize 48 category-specific training loops for a product propensity model previously orchestrated in Kubeflow. Alligned training to leverage distributed Ray clusters for resource elasticity, enabling efficient hardware utilization across workloads. Ruduced training runtime and infrastructure costs by migrating from Vertex AI custom jobs to KubeRay on GKE, cutting compute spend by ~40%. Automated Patient Record Synthesis via Vertex AI Pipelines. Implemented a cost effective and scalable batch prediction system using Vertex AI Pipeline to process high volumes of clinical documentation and generate automated patient summaries for Nursing staff. Implemented a structured document-to-folder workflow that automated the ingestion of diagnostics and prescriptions, utilizing Python and GCP services to maintain organized, HIPAA-compliant data hierarchies.

Automated retraining & Monitoring for object detection model in Autonomous Driving System. Designed retraining pipelines that bench new data against existing models, storing Model-IDs and evaluation metrics in BigQuery for tracking. Implemented automated alerts when retrained models outperform production, enabling data-scientists to review and approve deployment to production. Built continuous monitoring for data drift and skew by instrumenting inference pipelines with structured logging in cloud logging, and then visualizing derived metrics in Grafana Dashboard. ML Model Training/Scoring Data Pipeline. Designed and implemented a robust data pipeline for a machine learning project, ensuring efficient data flow from the source database to the data-warehouse. Established a Kafka-based synchronization mechanism to maintain real-time consistency between the source database and the data warehouse. Multi-Channel Pipeline for Dashboard refresh and Disaster Recovery. Created schema and data model to push data from spark to sql server using pyspark jdbc connector. Established synchronous schema on sql-server to ingest incoming streams of data from pyspark ETL scripts. Stored procedure to have backup of DB on remote server. Model High Performance Testing (HPT). Created an HPT framework to keep checking model health for Data Drift and Model Drift. Patent accepted for this novel HPT project.

ETL migration from SAS to Spark. Migrated a sas based ETL for process entire US healthcare benefits data from SAS based ETL to pyspark. Built an ROI simulator to use Market Mix Model output and elasticities to predict impact on sales by changing & promotional spends. Tableau based Healthcare dashboards for Healthcare providers across US. Automated dashboard refresh through Tableau server command-line-utility for any pre-defined cadence.