Own and lead engineering for DynamoGuard, the company’s primary real-time moderation product, ensuring reliability, delivering features, and driving cross-team coordination

• Built end-to-end pipeline for online model training and deployment, enabling customers to train and serve guardrail models self-serve and scaling the platform to more customers

• Led engineering initiatives to scale and harden systems, consolidating deployment images (7 → 3) to reduce complexity and spearheading a MongoDB → Postgres migration to address scalability bottlenecks

• Drove GPU slicing for model inference, cutting inference cost 10× for low/medium-throughput workloads through PoC benchmarking and deployment configuration optimization across SaaS and PaaS

Thesis: "Characterizing and Optimizing Networking Stack in Databases" advised by Michael Stonebraker

• Profiled VoltDB networking stack using perf, identifying server-side processing as a major bottleneck (~15% of query latency), exceeding execution time on transactional workloads

• Engineered kernel-bypass networking layer using DPDK and F-Stack via Java Native Interface (JNI), replacing the kernel TCP/IP stack and reducing server network overhead by up to 80%

• Improved database performance, achieving 44% higher throughput and 28% lower latency on Retwis, and 2× throughput on TPC-C at low concurrency.

Assisting students in understanding course materials. This involves helping them think through labs presenting applications of machine learning concepts and holding office hours to help with concepts and homework.

• Developed the backend for a new product offering in NestJS and MongoDB to let clients evaluate private data leakage in their large language models (LLMs)

• Designed the privacy evaluation pipeline to reduce the cost of inference by using AWS Sagemaker

• Added support for privacy evaluation on custom trained GPT models from OpenAI

• Built an automation tool to enable customers to create Jira tickets of security vulnerabilities directly from the Balbix dashboard with single click- down from more than 10 clicks (around 3 minutes)

• Implemented an ingestion pipeline for customer's security data from 3rd party source into central PostgreSQL database asynchronously via an Apache Kafka-based system