At Navi Technologies Ltd., I designed and shipped an LLM-based on-call agent that correlates logs, metrics, GitHub changes, and CI/CD state to automate first-line triage and root-cause analysis. It cut on-call MTTR from 14 hours to 2 and raised SOP coverage from 20% to 99%.
I also architected a clickstream and attribute feature store for real-time ML inference, with Spark batch and Kafka streaming ingestion processing 1TB+ daily. Its Java feature services served 60K peak RPM at p99 latency under 20ms.
At Navi, I led cross-team cloud cost initiatives that saved $67K/year, including service migration from AWS to OCI and automated Databricks and Kubernetes right-sizing. I also built automated monitoring for Data Science infrastructure, including model drift and data-source freshness checks.
Earlier at Navi, I developed a UPI feature store with Apache Flink ingestion and a Scala feature computation library, and automated model training workflows that cut training time by 50%. At American Express, I trained a BERT model for merchant embeddings and built its data-processing and visualization pipeline; at KAUST, I implemented GAN image inversion and evaluated StyleGAN2-ADA embeddings.

