Boston, MA.

• Designed and built a computer-vision-based testing pipeline in Rust for Amazon’s next-generation dashcart, connecting edge devices to AWS IoT Core over MQTT, using gRPC for inter-module communication, and resolving credential provisioning for newly onboarded cart units via AWS Lambda.
• Built and deployed an automated agentic testing workflow on AWS AgentCore that reads the cart’s on-screen display to flag anomalies during rollout and streams results to a real-time monitoring pipeline, with the goal of removing the need for on-site engineers during beta provisioning on the test fleet.

Beijing, China.

• Contributed to the pre-release FluxVLA Engine, an end-to-end VLA engineering platform spanning data processing, fine-tuning, evaluation and real-robot deployment across 7 mainstream VLA models (OpenVLA, GR00T, π0/π0.5 and more) on Franka, UR3 and ALOHA hardware; now open-sourced (500+ stars, 60+ forks) and deployed on Alibaba Cloud to serve VLA training and inference.
• Built a unified codebase for several VLA models including OpenVLA and π0, resolving environment inconsistencies between them and transplanting OpenVLA-OFT onto the LIBERO benchmark.

• Built an ARX robotic-arm teleoperation data collection, training and evaluation pipeline to replicate results from Diffusion Policy (DP) and Flow Matching on basic and contact-demanding wiping and flipping tasks.
• Incorporated joint torque modality into raw DP, observing improved action performance on contact-rich tasks.

• Taught Python-based computational problem solving; ran weekly labs and discussion sections on programming exercises, debugging and algorithm design.
• Mentored students with little or no prior programming background through office hours, applying programming to data analysis, modelling and simulation.

• Co-authored the CVPR 2024 paper MovieChat (700+ citations), a widely adopted benchmark for multimodal LLMs and vision-language models in long-form video understanding requiring video memory.
• Built the data collection software stack and annotated a high-quality benchmark with Q&A pairs and 100-word summaries for 1000+ ten-thousand-frame videos, released on the Hugging Face Hub.

• Led weekly discussion sections on Boolean algebra, sequential logic and finite state machines, reinforcing theory with applied problem solving.
• Supervised laboratory sessions on circuit simulation, C programming and system-level design, covering hardware–software integration and LC-3 assembly.