ML Jobs Infra

Designed and implemented K8s data plane and networking infrastructure for OCI’s multi-tenant AI training platform.

Implemented OSS mounts in Kubernetes using Rclone as a sidecar and custom FUSE mounts for ML workloads to support checkpointing of their distributed batch training jobs.

Built custom multi-homed networking for large-scale training and inference workloads using Multus CNI and contributed to the development of a custom IPAM solution to enable scalable, cloud-native pod networking.

ML Pipelines

Tech lead for ML Pipelines and shipped multiple features since it's GA.

Designed and implemented a novel solution for our serverless spark service’s integration with ML Pipelines using Kafka events and a cross tenant rule which protects event delivery, which will allow customers to make their data storage of model artifacts transient using a set of ephemeral tokens which bolstered Oracle’s stand on data protection.

Architected a multi-region distributed integration testing framework acting as a global canary system, proactively detecting failures and improving reliability across OCI realms.

Key contributor to the GA launch of ML Pipelines on OCI, enabling end-to-end machine learning workflow orchestration for large-scale distributed training and inference; authored the official Terraform provider (Golang) to support infrastructure-as-code automation.

Co-architected a real-time streaming model inference platform (“Stream Manager”) with a Java control plane and Golang binary inference runtime, delivering scalable, production-grade model serving using co-located containers with internal streaming endpoints safeguarded by linux network namespaces.

* Redesigned the backend architecture of a micro-service to work with larger network payloads using Python, Flask and RabbitMQ.

• Built a Proof of Concept to read, parse and execute existing Terraform scripts.

• Built a monitoring tool using Apache Spark and Python to analyze issues with internal network modules in Ansible and cut development time by 80%.

* Worked with VMWare to create an SDN automation using nodeJs of their BGP protocol routing onboarding for internal customers.

* Built an ETL data pipeline to process 3 Million employer names, created an algorithm to deduplicate it and created attributes with it.

* Used RSpark and Sparklyr to build a model to predict the default rate for a given loan amount and visualized it using ggplot.

* Reduced time taken to identify industry segments and its default rate, given an employer, by 150% and presented to key stakeholders

* Created a PoC using Azure IoT, DataBricks and PowerBI to detect and predict sensor variations in the factory

* Created dashboards to generate and visualize customer metrics in PowerBI connected to SAP HANA.