Suyeol Yun

Suyeol Yun is Head of AI at LinqAlpha, where he builds evaluation pipelines and uses their feedback to guide coding agents and the team’s development work. His career has moved from facial-animation engineering through political-science research to financial AI. His projects include an open dataset of World Trade Organization disputes, a research assistant for studying Congress, and systems for extracting financial data and evaluating investment-research agents.

From facial animation to political institutions

Yun studied mathematics as an undergraduate and spent his first four to five years as a deep-learning engineer working primarily in computer vision. His facial-reenactment work involved reconstructing three-dimensional meshes from hundreds of photographs taken on a lighting stage, then designing a graph neural network to transfer expressions from one person’s sequence of meshes to another’s. The job combined graphics programming in C++ with experimental decisions about neural-network architecture and training. His path subsequently took him through the U.S. lobbying industry and into doctoral study in political science at MIT. He left the doctoral program without completing a PhD; MIT lists him among its fall 2023/spring 2024 graduates with an SM in political science.

His DeepWTO project gave researchers shared data and code for studying legal language in WTO disputes. Each example pairs a description of a government measure with the text of a legal provision and a label indicating whether that provision was cited. Preparation code and baseline neural classifiers let researchers reproduce the task and test other approaches. Yun also identified an important evaluation trap: because most candidate provisions were not cited, a classifier could achieve high accuracy simply by predicting the majority label. The repository therefore emphasizes AUC-ROC rather than raw accuracy. DeepWTO earned him one of the inaugural MIT Prizes for Open Data in 2022.

Political institutions remained part of his research as language-model agents emerged. With Joseph R. Loffredo, Yun coauthored research introducing CongressRA, an agent designed to support scholars studying the U.S. Congress. The approach connects language models to external knowledge and specialized tools for collecting data, processing documents, and conducting analysis. It explores how those capabilities could reduce the work required to replicate, test, and extend empirical studies.

Building financial research agents—and ways to test them

At LinqAlpha, Yun’s collaborative engineering and research address different stages of an agent’s work: using tools, turning documents into queryable data, finding relevant evidence, and judging the resulting research.

• Recovering from tool failures. Yun coauthored the open-source DeepSeek-R1 function-calling agent with Subeen Pang, Yongjin Kim, and Chanyeol Jacob Choi. The framework combines iterative model calls, parallel tool execution, and JSON validated with Pydantic. When a tool fails or a response cannot be parsed, the error returns to the model as feedback, allowing it to revise its output or function arguments and try again.

• Making financial documents queryable. He coauthored Structuring the Unstructured, which proposes two specialized agents for extracting and querying financial key performance indicators and guidance. An extraction agent identifies quantitative information, standardizes its format, and checks its accuracy. A text-to-SQL agent translates natural-language questions into executable database queries. This separates preparing structured financial data from answering questions against it, without requiring users to know the database schema.

• Separating retrieval decisions. Yun is a coauthor of FinAgentBench, a benchmark containing roughly 26,000 expert-annotated examples concerning S&P 500 companies. It evaluates two decisions independently: selecting the relevant document type and locating the key passage within the selected document. That separation makes it possible to measure where retrieval succeeds or fails, rather than treating the entire process as a single result.

• Scaling expert-guided evaluation. His coauthored 2026 preprint FinAutoRubric tackles the cost of writing detailed grading criteria for financial research. Experts supply reusable guidance; agents and code generate, review, and validate task-specific rubrics. A writer agent researches expected values as of an information cutoff, a reviewer checks them, and unresolved failures escalate to a person. The accompanying benchmark contains 100 queries across 78 tasks and eight asset classes.

Yun also treats his role as helping colleagues reason about uncertainty. In his account of joining LinqAlpha, he describes debate and strategic framing as part of the job alongside engineering. Evaluation supplies feedback about whether the system is improving; his broader aim is to help the team articulate what it does not yet know and adjust its decisions as the evidence changes.