Diyuan Deng

• With the testing team on AppGallery ArkTS development: optimized the SDD workflow through end-to-end self-verification; packaged Hypium-based environment setup, test-script execution, and result capture as reusable skills, replacing manual compile–deploy–verify steps in DevEco Studio.
• Introduced AI into the closed loop: routed test failures to code agents for automated repair and integrated the skill chain with SDD compile/repair automation, closing the code–test–feedback–fix pipeline from test-department scripts to developer-ready fixes.

• Optimized throughput and distributed communication for enterprise-scale Qwen3-class LLMs (4B / 30B) using SGLang and dual-container isolation; built a load-test web platform; two patent applications in progress from core optimizations.
• Configured and benchmarked parallelism strategies (PD/EPD, TP/DP/EP) and FP8/INT8 quantization across RTX 4090 and H20 clusters; achieved TTFT −48.9% and throughput +664% on Qwen-30B (TP4+DP2), and throughput +750% on Qwen-4B (TP1+DP2).

Collaborated with Prof. Bo Zhao's group (Shanghai Jiao Tong University) on embodied AI data and VLA models.

• Built the xArm data pipeline and optimized VLA (8,000+ trajectories).
• LeRobot stack with multimodal LLMs and natural-language command interfaces; grasping and multi-layer stacking tasks at >90% success.