Built a dynamic model routing and post-processing system for Amazon's latest LLM model, enabling adaptive “Agent Thinking” based on query complexity.

Helped improved helpfulness metrics from 26.8% to 37%+ by supporting flexible transitions between thinking, planning, and acting modes.

Helped shift from rigid architectures to a unified ReAct framework with PPO-based reinforcement learning, reducing reliance on IFT/DPO.

Delivered a model-agnostic RAG system with content attribution, supporting integration with 60+ premium publishers.

Proposed and implemented distributed post-processing architecture to simplify scaling and system complexity.

Helped enabled cross-LLM web search with transparent and high quaility third-party sourcing

Internationalized core retrieval service to India and UK within 3 months of US launch.

Automated cross-region setup, reducing infrastructure build time by ~70%.

Built region-aware infrastructure processing 20M+ queries in India and 5.8M in the UK with consistent performance.

Migrated workloads from p4de to g5 and p3 GPU instances, cutting infrastructure costs by ~90%.

Localized GPU infrastructure to eliminate cross-region latency and improve offline indexing throughput.

Enabled international launch by resolving cross-region issues, freeing up shared resources for critical tasks like model training.

Designed a cache-aside system across Orchestrator, Retrieval, Model Server, Cache to support post-purchase Q&A.

Built centralized signal routing logic and content-aware filtering to improve modularity and response quality.

Improved post-purchase coverage by 5%, with full monitoring and alerting for reliability at scale.

Developed an event validator for customers to validate event schema and SQL expression before data on-boarding

Solved complicated events support request by more than 10 Alexa shopping data producer teams

Proposed complex-value convolutional neural network to determine the frequency of blur in magnetic resonance image

Implemented generative adversarial networks to generate field map for MR image deblurring and got 6% improvement on accuracy

Implemented a pipeline with parallel computing combining object segmentation and visual representation learning for medical image classication and improved F-score by 5% in clinical disease prediction(Pytorch)

Proposed machine learning methods for medical image classication, object detection and segmentation

Reviewed over 100 recent papers of machine learning in medical image analysis and reproduced the core methods for comparison and discussion (Pytorch, Caffe, Tensorflow)