Research · GPU Systems

• Synthesized and optimized high-performance CUDA/Triton GPU kernels, achieving a 2.4x speedup over optimized PyTorch baselines by engineering an execution-grounded Reinforcement Learning pipeline utilizing Qwen2.5-Coder
• Elevated the KernelBench fast@1 success rate from 14% to 58% by fine-tuning the solver via GRPO, leveraging a correctness-gated reward function that mapped raw compilation logs and runtime latency into a dense learning signal
• Automated parallel RL environment rollouts and scaled training infrastructure, managing a multi-node, 16-GPU NVIDIA H200 cluster by architecting a distributed parallel-rollout and trace-collection harness built on SLURM, Docker, and Python

Startup · LLM Infra

• Ensured high-availability document parsing for 4K+ monthly uploads, tracking execution latency and LLM trace metrics via Langfuse, by architecting an asynchronous verification pipeline using Gemini VLM, FastAPI, and Celery
• Reduced database read latency by 10x while maintaining strict company-scoped RBAC, validated through query execution benchmarks, by migrating relational queries to precomputed JSONB snapshot tables with event-driven incremental updates
• Scaled B2B billing operations to process $500K+/month in recurring revenue by designing and implementing Stripe-backed subscription infrastructure with complex proration logic

Enterprise · Mobile

• Engineered a Spring Boot Backend-for-Frontend (BFF) layer processing 2M+ daily requests, cutting P95 latency by 65% via Redis caching with a tag-based, Kafka-driven invalidation pipeline to guarantee mobile read consistency
• Cut mobile notification latency by 40% for 50K+ users, processing 1M+ daily events with guaranteed reliability, by building an event-driven Kafka pipeline featuring exactly-once delivery semantics, structured dead-letter queues, and batched FCM delivery
• Delivered a mobile platform handling 3M+ daily work orders, earning the CBRE CDTO Excellence Award, by serving as the primary engineer architecting offline-first synchronization via MongoDB Realm and exposing unified workflows with GraphQL

Enterprise · Data Platform

• Reduced global reporting latency from 24 hours to under 5 minutes while eliminating extraction failures across 12 downstream integrations by engineering near-real-time Change Data Capture (CDC) pipelines using Debezium and Snowpipe
• Automated global analytics pipelines across EMEA, APAC, and US business units, consolidating data flows from over 20 microservices, by architecting dimensional star schemas in Snowflake and orchestrating transformations via dbt and Airflow

Enterprise · Platform

• Maintained high-throughput validation across 2M+ daily maintenance requests with zero-latency overhead by building a request-flagging microservice using Spring Boot, Hibernate, and PostgreSQL to execute real-time business rule checks

Enterprise · Mobile Platform

• Led a 4-developer intern team to build an employee self-service portal within FPT Software's mobile app, serving 30K+ employees, utilizing React Native, Redux, and GraphQL on the front end, and Node.js, Express, MongoDB, and AWS S3 for backend features
• Engineered CI/CD pipelines for mobile deployments with Azure DevOps, automating builds and deployments to app stores

Enterprise · Automation

• Automated invoice processing, saving 10+ hours per week by building Selenium & Python web crawlers to scrape XML invoice data
• Containerized and deployed crawlers as microservices with Docker & Kubernetes, improving scalability and deployment efficiency