Skip to content
View chi030303's full-sized avatar
💭
learing
💭
learing
8 followers · 14 following

Achievements

Achievement: Pull Sharkx2

Achievements

Achievement: Pull Sharkx2

Block or report chi030303

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. Markdown is supported. 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
chi030303/README.md

Typing SVG

M.S. in Applied AI, Nanyang Technological University  ·  LLM Algorithm Intern, Huawei
shock22330@gmail.com  ·  LinkedIn  ·  Available for full-time roles from late Oct 2026

I work on LLM post-training: distributed RL (GRPO / GSPO / DPO), agentic pipelines, and the systems work that makes them run on large GPU/NPU clusters. Previously I built production Agent/RAG services and code-LLM data pipelines. I care about the path from alignment algorithms to stable, measurable training.

My Skill Tree 🌳

Post-Training & RL Languages Systems & Serving
GRPO GSPO DPO RLHF SFT PyTorch Python Go Java C++ vLLM Ray Megatron AReaL
Agents & Data DevOps & Cloud-Native Backend
LangChain LangGraph LangSmith RAG Airflow Docker K8S Jenkins Git FastAPI Kafka ES

Main Quests 🛡️ (Professional Experience)

Huawei LLM Algorithm Intern @ Huawei · partnered with HiSilicon on Ascend NPU post-training
Feb 2026 – Aug 2026
  • Agentic RL: built post-training pipelines for Qwen2.5-Coder 7B with vLLM + Ray on a 32-NPU cluster; integrated TIS / CISPO / GSPO / TBPO. +10% AIME24 with no drop on AIME25.
  • Disaggregated MoE RL: GRPO for Qwen3-30B-A3B on a 48-NPU cluster with AReaL + Megatron-LM; decoupled resource allocation cut per-iteration time by 66%.
  • Multimodal RL (T2V): online GRPO for HunyuanVideo-13B on 32 NPUs; CPU/NPU offloading for the DiT, text encoders, and VideoAlign reward; 60% faster iterations after communication tuning.
  • Low-bit pre-training: 8-bit / 4-bit OLMo-1B on 300B tokens (128 NPUs); Online Softmax to avoid NPU overflow; reported PPL / MMLU / HellaSwag / ARC via lm_eval.
RockFlow Python Development Intern @ RockFlow
Oct 2024 – Mar 2025
  • Multi-agent workflows with LangChain / LangGraph for investment Q&A and natural-language trading.
  • Streaming content pipeline with Kafka + FastAPI; Agent eval on LangSmith (A/B + Judge LLM), ~90% regression coverage.
  • Airflow DAGs for scheduled BI / Feishu reporting (~20 person-hours / month saved).
Zhipu AI Python Intern @ Zhipu AI (智谱AI) · ChatGLM code-data pipeline
Mar 2024 – Jun 2024
  • Code-generation and bug-fix datasets for ChatGLM: guidelines, validation scripts, and sampling QA on logical / functional correctness.

Side Quests & Achievements 🏆

  • LMSYS Chatbot Arena Human Preference Predictions (Kaggle)
    • QLoRA SFT of Qwen3-14B + DPO on non-tie pairs (RTX 6000); stacked LLM OOF with GBDT. Log loss 0.95 (competition winner 0.96).
  • 📜 "Network Security Threat Detection System Based on Knowledge Graph"
    • Second author; YAC 2025 / IEEE Xplore. Multi-source KG (ATT&CK, CVE, D3FEND) on Neo4j; RAG with Weaviate + GPT-4o / Llama 3 routing.

Leveling Up 🎓

M.S. in Applied Artificial Intelligence @ Nanyang Technological University (Nov 2025 – Oct 2026)
B.E. in Software Engineering @ Beijing Jiaotong University (Sep 2021 – Jun 2025)

Player Stats 📊

GitHub Stats

Top Languages

GitHub Streak

Contribution Snake

Profile Views

Pinned Loading

  1. MoviesRecommendationSystem MoviesRecommendationSystem Public

    Java 1

  2. F1yingWhite/API-Big-Work F1yingWhite/API-Big-Work Public archive

    Vue 2

  3. Intelligent-elderly-care/algorithm Intelligent-elderly-care/algorithm Public

    算法模块

    Python

  4. QigeOS QigeOS Public

    Experiments of Operating system course.

    Makefile

  5. Wechat_golang Wechat_golang Public

    利用golang编写的微信公众号后台系统

    Go

  6. moviesDataAnalysis moviesDataAnalysis Public

    对kaggle上的电影数据集进行特征分析和特征融合

    Python 1