Skip to content
View wangkai9161's full-sized avatar

Block or report wangkai9161

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
wangkai9161/README.md

王晓凯 | Xiaokai Wang

哈尔滨工业大学应用数学硕士研究生,关注 LLM Inference、搜广推、LLM Agent/RAG 与自监督学习。我希望把可复现实验、性能测量和实际科研问题连接起来,而不是只展示模型名称或框架调用。

Focus Areas

Direction Repository What is verified
AI Infra / LLM Inference ai-infra-lab 流式推理压测、TTFT/TPOT/ITL、并发与长度矩阵、CUDA Graph 和 profiling 实验
Search, Ads, Recommendation search-ads-recommendation Criteo CVR、双塔召回、负采样、序列推荐、排序与统一离线评估
LLM Agent / RAG llm-agent-lab 工具调用、参数校验、日志诊断、混合检索、来源引用、拒答与自动化测试
Machine Learning Research ml-research-lab 自监督学习、弱标签训练、地震混采数据分离与信号恢复研究

Current Engineering Evidence

  • 在单张 RTX 5080 上搭建 vLLM 推理实验,分析并发扩大时吞吐与 P95 TTFT 的权衡,并对 eager/CUDA Graph 路径进行对照。
  • 基于 Criteo Sponsored Search 约 1,600 万条点击日志开展点击后 CVR 实验,比较 LR、FM、Wide & Deep、DeepFM,并使用时间切分、校准和消融约束结论。
  • 基于 MovieLens-1M 与 LastFM 风格隐式反馈任务实践双塔召回、BCE/BPR、GRU4Rec、SASRec、DIN、DeepFM 和多兴趣/生成式推荐。
  • 构建面向科研实验的 Tool Agent 与可追溯 RAG 原型,覆盖 GPU 检查、训练命令生成、日志诊断、来源引用和无依据拒答。

Research

  • Conference paper: Reblend-denoising inspired self-supervised strategy for inversion-based seismic deblending.
  • Journal submission: Flexible reblend-denoising inspired self-supervised iterative seismic deblending with a hybrid norm loss function.
  • Two invention patent applications related to self-supervised iterative deblending and interpolation reconstruction.

Technical Stack

Python · PyTorch · NumPy/Pandas · Linux · Git · FastAPI · vLLM · Transformer · Recommendation Systems · Self-supervised Learning

Working Principles

  • Reproducible configurations and measured evidence before performance claims.
  • Explicit dataset, split, workload, baseline, and metric definitions.
  • No credentials, private datasets, model caches, or unverifiable production claims in public repositories.

Popular repositories Loading

  1. search-ads-recommendation search-ads-recommendation Public

    Reproducible search, advertising CVR, retrieval, ranking, and sequential recommendation experiments.

    Python 1

  2. wangkai9161 wangkai9161 Public

    Xiaokai Wang - AI Infra, recommender systems, LLM agents, and self-supervised learning.

  3. ai-infra-lab ai-infra-lab Public

    LLM inference serving benchmarks, profiling, and reproducible AI infrastructure experiments.

    Python

  4. llm-agent-lab llm-agent-lab Public

    Tool-calling and retrieval-augmented LLM applications with tests, evaluation, and guarded execution.

    Python

  5. ml-research-lab ml-research-lab Public

    Self-supervised learning, model training, and signal restoration research projects.