Leading LLM infrastructure development and optimization

• Extended in-house LLM context window from 32k to 180k tokens using advanced distributed training.
• Optimized LLM inference service achieving 6.7x throughput improvement through system-level optimization.
• Built production evaluation pipeline processing millions of tokens for in-house LLM deployment validation.
• Created model artifact interface reducing deployment cycle from months to days.
• Reduced OpenAI API costs by 35% through prompt engineering optimization.

Building AI solutions for game development

• Optimized MLOps pipeline for Microsoft Recommenders, reducing runtime by 30% and cost by 20%.
• Built Azure infrastructure for distributed training of game navigation AI agents.
• Developed gRPC framework enabling RL agents to interact with Unreal Engine games.

As a ML engineer in the Microsoft AI Development Acceleration Program (MAIDAP), I built AI solutions for different product domains, such as News, Cognitive Service, IoT, and Data Analytic Platforms.

• Developed an open-sourced tool, genalog, to generate synthetic documents that improved Named Entity Recognition (NER) model performance by 76% on scanned documents. Publication
• Designed and built an end-to-end machine learning solution for ad personalization on Microsoft News that showed 8% increase in ad revenue.
• Designed and implemented CI/CD infrastructure for ML model deployment in a custom Spark Cluster, a critical test framework for the in-database inference features in development.
• Developed a reinforcement learning-based HVAC control system, establishing data pipelines for sensor data aggregation and online model inference.