Austin, TX
Po-Chen (Brandon) Yeh
Building with data, thinking about AI, and taking pictures along the way.
About
I’m a master’s student in Information Studies at The University of Texas at Austin. Working in software, I’ve seen how powerful AI is — and how unsettling it can be — so I keep coming back to one question: how do you let AI do more without people losing track of what it’s doing?
Before that, I spent a year and a half building manufacturing data pipelines and dashboards as a contractor for Google, and earlier worked on backend data infrastructure at Appier. I studied Information Management at National Taiwan University, then came to Austin because I’ve always wanted to see more of the world — and find out what else I might be capable of.
Outside of code, I take photos and video on a Fujifilm X-T5. Time seems to move faster every year, and a photograph is the closest thing I know to pressing pause — a way to hold on to the good people, places, and moments I come across. I also lift weights, travel, and read.
Photography
See all photos →
Projects
Cost- and Privacy-Aware LLM Router
2026 – nowUT Austin · Post-Training of LLMs · Team projectWhen developers ask an AI tool for help, secrets slip in without anyone noticing — an API key buried in a config file, someone’s email address in a log snippet. Sending every request to the biggest external model is expensive, and sending the wrong one quietly leaks those secrets.
We’re post-training a small, internally hosted model that sits in front of the others. It reads each request, spots credentials and personal information, and decides where it should go: a cheap external model, a strong one, or a model that never leaves the building. The twist is in the reward — a single leak costs more than any amount of money saved.
Governable AI Agents for Shared Research Knowledge
2026 – nowUT Austin · Human-AI Interaction · Team projectResearch labs lose knowledge every time someone graduates — it’s scattered across tools, half-documented, or never written down at all. AI agents could help keep a shared lab wiki alive, but once an agent can edit what everyone relies on, a new question appears: what should it be allowed to change, and who gets the final say?
We’re interviewing researchers about what they’d actually hand off to an agent, and designing a review flow where every AI-proposed edit arrives with its sources, rationale, and how confident it is — closer to a pull request than a chatbot answer. I’m working on the user research and the system architecture, including a prototype of how an agent searches a lab’s documents and turns what it finds into a traceable proposal.
Previously
- Intelligent Manpower Corp.BI Engineer, contracted to Google2024 – 2026
- AppierBackend Engineering Intern2023 – 2024
- IBM ConsultingApplication Consultant Intern2023