Hi, my name is

Mingxuan (Kevin) Nie.

I make LLM output reliable enough to ship.

I'm an MS CS student at UC San Diego, and I build LLM agent systems: multi-agent loops, tool calling, MCP, and the grounding, validation and repair around them that turn a model's fluent, plausible, wrong output into something you can run.

Most recently I built an agent that turns QA test cases into on-device test scripts at miHoYo. Before that: a multi-agent image generation and storyboarding system at Vicino AI, an LLM on-call agent used by 1,000+ engineers at TikTok, and Android network diagnostics for Lazada at Alibaba.

Experience

Playground

RogueForge is a roguelike deck-builder where an LLM forges every card, enemy, event and hero during the run. The model never touches game logic; it picks from whitelisted effects, and this validator decides whether the result is fair before a player sees it. It is the real thing, ported line for line from the Python backend.

A card's power is a weighted sum of its effects. The acceptable band for each (rarity, cost) bucket comes from the real Slay the Spire Ironclad cards in that bucket, so "balanced" is defined by data, not by taste. In band: accepted. Slightly out: clamped one point at a time and the rules text re-rendered. Far out, or structurally wrong: rejected, with the reason sent back to the model.

Try 40 damage on a 1-cost card, poison on an Attack, or a Rare that gains 2 Block.

2,658candidates logged: 86% accepted, 9% repaired, 5% rejected
87%of real Ironclad cards fall inside their own band

Projects

Skills

Education

What's next?

Get In Touch

I'm looking for a full-time software engineering role starting spring 2027. AI agents, LLM infrastructure and evaluation are where I have the most to offer, but I care more about shipping things that work than about the label. If you are building something like that, I would like to hear about it.

Say Hello →