LLM / VLM engineer · Computational science background

Walker Chi 迟明远

Building foundation models,
from data to training.

I build the data pipelines, training workflows, and evaluations behind large language and multimodal models. Previously at Moonshot AI and MiniMax; MSc in Computational Science and Engineering from ETH Zürich.

Beyond pretraining, I keep coming back to space elevators, fusion energy, and how we understand and simulate the physical world.

Download CV ↓GitHub ↗LinkedIn ↗Email ↗

01 / Professional experience

Work & contributions

Full CV ↗

Apr – Sep 2026
Beijing, China

Moonshot AI

Scaling LLM Researcher

Pretraining data. Led the end-to-end pipeline across acquisition, crawling, procurement, parsing, and processing; ran multi-epoch and scaling-ladder experiments.
Agent evaluation. Contributed seed data, tasks, and agent trajectories to Agent Last Exam (ALE), evaluating planning, reasoning, and execution.
GPU compiler capability. Independently built a MiniTriton case study to evaluate kernel generation, correctness, and optimization.

Pretraining · Scaling experiments · Agent evaluation · GPU kernels

Sep 2024 – Apr 2026
Shanghai, China

MiniMax

VLM Researcher

M-series model development. Contributed across LLM / VLM data construction, pretraining experiments, hyperparameter search, evaluation, and ablations.
Multimodal data at scale. Owned data sourcing, processing, quality control, and delivery. Built TB-to-PB-scale pipelines producing 20T+ interleaved multimodal tokens.
Quality & infrastructure. Developed data reward / quality models and ablation workflows; implemented multimodal DPO in Megatron-LM and improved data packing and evaluation scheduling.

LLM / VLM · Data quality modeling · Megatron-LM · Distributed pipelines

Jul – Sep 2021
Shenzhen, China

Tencent

Algorithm Intern · TEG Information Security

Graph neural network optimization. Applied sparse computation and JIT compilation to GEN (Graph Estimation Neural Network).
Measured improvement. Reduced memory usage to approximately 1/1,000 of the original implementation and accelerated inference by approximately 15×.

Graph neural networks · Sparse computation · JIT

02 / Education