Research grounded in building.

• Technical owner of Apodex Harness infrastructure.
• Built workflows for long-running work with parallel agents, cross-branch synthesis, global verification, and continuation after interruption.
• Proposed schema-constrained, planner-generated workflow DAGs and built a compact execution engine with branching and verification-retry loops.
• Developed token-budgeted context compression, file-based research memory, and evidence checks for unsupported numbers and citations.
• Improved sandbox isolation, tool permissions, durable agent messaging, and cancellation-safe event streaming; connected workflows to unified evaluation and model-release pipelines.

• Led tool-use post-training for Watt’s Llama 3.3-70B model, which reached No. 1 on Berkeley’s Function-Calling Leaderboard V3 in December 2024.
• Built workflows supporting more than 200 tools and improved multi-step task completion from 70% to 90% in a 1,000-task evaluation.
• Evolved GMate into a closed-loop marketing agent with preference, factual, and experiential memory; led launch reliability for the Xiaohongshu product in December 2025.
• Architected the TikTok Shop product-selection assistant with LangGraph, DOM-and-vision browser automation, observability, and retrieval-enhanced evaluation.
• Built hybrid retrieval over a 6,000-document recruiting knowledge base and a 500,000-candidate talent graph for W-Hiring.

Responsibilities include:

• Tech leader of a team of 7 AI and data engineers on the development and deployment of robot navigation and IoT analysis.
• Lead development of “Multi-embedding query for person MOT-ReID System”, which uses the innovative idea of pose-guided feature alignment to improve accuracy.
• Development of human action recognition and intrusion detection, with the combination of spatio-temporal and skeleton-based models. The algorithm has been deployed in applications like construction site surveillance and facility management.
• Construct an object detection/segmentation pipeline with quantization-aware mixed precision training on top of DETR and Mask2former. The pipeline also supports semi-automatic labelling and curriculum learning.
• Development of model quantization and acceleration on edge devices of all video analysis related algorithms in both C++ and Python. The target edge computing chips include Rocketchip RK3588 NPU and Sophgo TPU.
• Development of long-term time-series forecasting and classification for stock return forecasting, as well as time-series data from IoT sensors for anomaly detection, remaining useful life prediction.

• Teach postgraduate EE7207 Neural Networks and Deep Learning, including attention, transformers, graph neural networks, and practical training and inference.

Thesis on Visual Metric and Semantic Localization for UGV. Supervised by Prof Mao Kezhi and Prof Han Wang. Published 6 conference papers and 4 journal papers.

GPA: 3.9/4.0

GPA: 3.4/4.0