Building the platform to run live LLM workflows in real time at scale. Designing architecture that supports fast iteration, enterprise reliability, and AI performance in production. Creating infrastructure that balances experimentation with operational rigor. Working closely with product teams to ensure AI features ship smoothly and deliver real business value.

Architecting for both latency-sensitive and high-throughput use cases. Enabling internal teams to move faster with safe and scalable abstractions.

• Led the architecture and production deployment of the first version of the AI Assistant for Acrobat, growing DAU from 0 to 20 million.

• Led the development of an AI Evaluation platform from the ground up, tailored for LLM-powered systems, enabling LLM evaluations, A/B experimentation, prompt tuning, and seamless CI/CD pipeline integration.

• Overhauled the Prompt Management System for Acrobat GenAI by introducing a real-time Prompt Store, enabling storing, versioning, tuning, and real-time loading of prompts in Stage/Prod environments.

• Spearheaded initiatives to improve team agility by introducing Model Routing and the Model Gateway services, increasing delivery efficiency by over 50%.

• Developed an access code mechanism at Adobe Sign for a large enterprise client (major bank), addressing a compliance issue in the APAC region by replacing URLs in signing emails, successfully enabling secure agreement signing and retaining a multi-million-dollar partnership.

• Led multiple impactful projects, features, and engineering initiatives, including Bulk Client Data Download and Document Link Expiration at Adobe Sign.

• Served as the Subject Matter Expert (SME) for Reporting and Apache Druid at Adobe Ad Cloud, driving petabyte-scale advertising data ingestion, aggregation, and querying.

• Adobe Ad Cloud (formerly Tubemogul)

• Responsible for building and maintaining the core data aggregation system that processes approximately 13 billion events daily.

• The architectural components of the system included Druid-powered real-time querying, Kafka streaming for aggregation, decoration, and storage, Hive/Presto (managed over Qubole) for internal business intelligence, data validation, scheduled/ad-hoc customer reporting, and testing, MySQL DB for storing customer account information, and Prometheus for metrics and alerting management. All of these components are hosted over Kubernetes and AWS.

• Served as the subject matter expert in Druid infrastructure and customer reports.

• Owned the Echo Spatial Perception (ESP) service, one of Alexa's core platform services that serves more than 100M requests a day and ensures single device response in multi-device environments

• Designed and developed data collection and storage solution for our ML requirements

• Optimized the device whitelisting/certification system from taking 2 weeks to an hour

• Designed and developed the testing framework for context-based device arbitration

• Worked on a project under Prof. Suman Banerjee to build an android app which rates a user’s driving on a scale of 10 based on the accelerometer and gyroscope data generated by the smartphone.