Mingxi Yan Carnegie Mellon · LTI

Some things I’ve built.
Some things I’m figuring out.

I build AI systems and the parts around them. I also play poker and other games, and sometimes invent new variants. I like philosophy, especially logic, and I write about love and growth. Usually, I want to know why something works the way it does.

Try them

Follow a thread

Things I spend time on

Open a note. See where it leads.

Engineering · TikTok01

Finding the right data is part of the problem.

An internal AI agent should be able to find a usable data source from a question. The user should not have to know the table name first.

Open the noteClose the note

During my data engineering internship, I worked on data discovery for internal AI agents and a revenue-diagnosis workflow for livestream operations. Both were validated with operations and business stakeholders and adopted into their working process.

What I worked on ↗
Research · CMU–Apple02

Who is asking changes the attack.

PersonaTeaming adds personas to automated red-teaming. I worked on how those personas are generated and evaluated the workflow across six target models.

Open the noteClose the note

I am the second author on this UIST 2026 paper. I suggested persona as an additional axis, worked on the dynamic generation mechanism, implemented the workflow, and wrote the technical results section.

Read about PersonaTeaming ↗
Course project · Poker + AI03

A good action. A convincing explanation. Are both right?

I worked on fine-tuning Llama 3 8B to produce a solver-backed poker action and explain it. Improving the explanation came with a small cost to action accuracy.

Open the noteClose the note

With Cassandra Shi and Harry Huang, I used solver-backed labels, teacher-generated explanations, and a rule-based verifier checked against 200 blind human judgments. The tradeoff between action accuracy and explanation quality is part of the result.

The project and paper ↗
Side project · Home games04

We wanted to play our own rules.

I like playing poker with friends and trying different rules. Dealer’s Choice became a small multiplayer platform for the variants we play at home games.

Open the noteClose the note

The platform supports custom variants and real-time play. I worked on the WebSocket connections, automated testing and deployment, and the AWS deployment. The rules themselves are on this site too.

Explore 12 poker variants ↗
Ongoing research · Sapling Lab05

When does someone start treating an AI as having a mind?

I am working on a simulation study of how these beliefs form and change over a long conversation.

Open the noteClose the note

The study looks at user characteristics, assistant social behavior, and conversational situation. I lead the NLP side and work on the simulation pipeline and turn-by-turn measurement. The manuscript is in preparation.

The ongoing study ↗
Poems & prose06

Sometimes I write to find out what I think.

Poems, an essay, and fragments. A different way of working through something that is not quite clear yet.

Open the noteClose the note

These are the pieces already collected on the personal side of this site. Some are in Chinese; some are in English.

Read 2 poems and 1 essay ↗

From our home games

Dealer’s choice.

12 variants we play at home games. Some are mine; some I picked up from friends.

One of mine

Auction

Economic variant where players bid to receive additional cards.

Read the rules ↗

A line I kept

“When we love, we always strive to become better than we are. When we strive to become better than we are, everything around us becomes better too.”
— Paulo Coelho, The Alchemist