Zhengqiang Yin银正强

Part of ByteDance’s AML team, working on LLM inference infrastructure spanning GPU clusters totaling over 10,000 GPUs.

• Contributed to building Ark’s LLM inference platform from inception to production as an early team member, supporting internal ByteDance workloads and external enterprise customers through 火山方舟 / BytePlus ModelArk.
• Contributed to distributed inference and performance optimizations, including prefill–decode disaggregation across heterogeneous hardware and persistent KV caching.
• Contributed to inference API capabilities, including reasoning-mode support and efficient constrained decoding for function calling and JSON Schema–constrained outputs, with minimal performance overhead.
• Previously worked on machine learning platforms, deep learning training acceleration, and domain-specific large language models.

Worked on search freshness with Bing's Fresh Core team.

• Built internal tools to support visual debugging of search systems.
• Investigated methods to identify freshness-sensitive queries in daily search logs.
• Explored reranking models trained with freshness labels.

Worked on automated review generation for iQIYI PaoPao with the Qiwen Tag team at IIG.

• Compared neural language models for review generation.
• Deployed a trained review generation model for the operations team.