Work Experience

[ Sep 2024 - Present ] Applied Scientist at Amazon, Applied AI - World-Wide Store, Bellevue, WA

Browser-Operating LLM Agent — Internal Automation Platform
Primary developer who took an early-stage agent codebase to production, owning the LLM interaction layer, DOM perception pipeline, agent runtime and evaluation harness.

  • Re-architected the agent harness around a frozen multi-turn history backed by long/short-term memory and a local filesystem, replacing a prompt that was rebuilt at every step; rebuilt DOM perception around semantic labeling, viewport-scoped element detection and a context-adaptive token budget.

  • Substantially raised the prompt-cache hit rate and cut per-step latency several-fold through forced tool-use structured output, cached schema computation and a reused non-blocking inference client.

  • Built the end-to-end evaluation harness (a resumable benchmark runner with an LLM-as-judge against human-annotated gold answers) and the 3-stage regression design used to gate production upgrades.

  • Result: higher task success than an open-source browser-agent baseline at a fraction of its cost and latency on an internal workflow benchmark; now deployed in production at large scale.

Autonomous Agent Debugging & Repair Pipeline

  • Built a staged LLM pipeline that ingests a production agent's specification and execution telemetry, attributes failures to specific workflow steps, and emits corrected, deployable agent configurations with no manual authoring.

  • Designed a multi-category production failure taxonomy with a category-conditioned remediation policy, reaching high root-cause attribution accuracy on a manually labeled benchmark of real production failures.

  • Applied to hundreds of customer agents escalated through production support, where most repaired agents improved customer-validated task completion while consuming fewer tokens than the original specification.

LLM Agent Evaluation & Reliability System

  • Shipped a reference-free LLM-as-judge system scoring production agent trajectories on instruction-following compliance and retrieved-information groundedness, resolving a 12-cell rubric into a 4-tier reliability verdict; now evaluating production agent traffic at scale.

  • Established the service's first quality metric (a healthy-run rate) as a release-gating regression signal, calibrated against a human-labeled reference set.

  • Designed the serving strategy (inline judging for user-triggered runs, sampled judging for batch traffic) to give fleet-wide coverage at a fraction of the inference spend, with failure-signal spikes auto-cutting on-call investigation tickets.

[ Jun–Sep 2021 & May–Sep 2023 ] Applied Scientist Intern at Amazon, Alexa AI, Sunnyvale, CA
Built an end-to-end audio-driven talking avatar pipeline: reconstructed photorealistic 3D head avatars from monocular video via NeRF (converging 40% faster than baselines through architecture and loss redesign), learned a vector-space expression representation for disentangled control over facial attributes, and drove avatars with speech signals for natural lip-sync and real-time head motion.

Education

[ 2018–2025 ] Ph.D. in Electrical and Computer Engineering, University of Minnesota, Twin Cities — dissertation: “Robustness and Safety of Deep Learning Models” (advisor: Ju Sun)
[ 2015–2017 ] M.S. in Electrical Engineering, Chalmers University of Technology, Sweden
[ 2011–2015 ] B.Eng. in Electrical Engineering, Shanghai Jiao Tong University, China

Professional Service

Reviewer

[ 2023–2025 ] Conference on Computer Vision and Pattern Recognition (CVPR)
[ 2023–2025 ] Transactions on Machine Learning Research (TMLR)
[ 2023 ] Neural Information Processing Systems (NeurIPS)
[ 2021 ] IEEE International Conference on Robotics and Automation (ICRA)
[ 2020 ] IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
[ 2020 ] IEEE International Conference on Robotics and Automation (ICRA)
[ 2020 ] IEEE Robotics and Automation Letters (RA-L)
[ 2019 ] IEEE International Conference on Robotics and Automation (ICRA)

Conference organizer

[ 2023 ] SDM23 assistant session chair.