Baixin Guo (Max)

I design and evaluate human-centered AI systems for evidence-bounded decision support. My academic background is in Applied Psychology, which informs how I define evaluation requirements and interpret system behavior. I am interested in cases where an AI output communicates more certainty than its evidence supports, and in interventions that address that mismatch without discarding useful information. CheckMyCoach is my primary research system. I co-designed its evaluation decomposition, specified system behavior, and reviewed outputs against target-removal and information-retention requirements. Human validation is in preparation; current evidence is limited to implementation and a bounded development run. InteractionKit and Knowledge Compiler support my methodological work through experiment specifications and evidence infrastructure. Separately, I contributed baseball data to collaborative injury-risk research with Professor Lorena Martin at USC Marshall's Trojan Sports Research Lab. That work produced two SABR abstracts, with related manuscripts in preparation.

My research asks how AI decision-support systems should respond when an output communicates more certainty than its evidence supports. I approach this as a systems and evaluation problem: define what is being checked, make the intervention inspectable, and distinguish an engineering result from evidence about human decisions. My Applied Psychology background informs the measurement questions, while system design makes those questions concrete.

CheckMyCoach is the center of this work. It routes selected outputs, assigns rule-based tags, generates candidate revisions, and checks both target removal and information retention. I co-designed the evaluation decomposition, specified system behavior, and reviewed outputs against target-removal and information-retention requirements. Its bounded development run exposed the difference between removing a target feature and satisfying every evaluation requirement. I am preparing human validation of these judgments. Human evaluation has not yet been conducted, so I do not claim improved trust calibration or decision quality.

Two supporting assets address different methodological needs. InteractionKit is released, frozen software for typed experiment specifications and contract checks; those checks do not establish construct validity. Knowledge Compiler supports typed evidence objects and partial provenance recovery; structural checks do not guarantee source fidelity. These assets are not evidence that an integrated platform has been validated.

Separate collaborative work at USC Marshall's Trojan Sports Research Lab gave me experience contributing data to baseball injury-risk studies. That experience belongs in my research record without changing the focus of my HCAI agenda. Looking ahead, I want to connect explicit decision costs, intervention policies, and baseline comparisons to human reliance outcomes. My aim is to make the relationship between a system action and the claim supported by its evaluation more precise, rather than treating a working prototype as the end of the research question.

I am seeking graduate research opportunities in human-centered AI and computer science for Fall 2027, particularly where system design and careful evaluation meet.

I design and evaluate human-centered AI systems for evidence-bounded decision support through system specification, inspectable evaluation structures, and evidence review.

View CV →

When I'm not calibrating LLMs, you'll find me —

Former student-athlete — I'm absolutely obsessed with basketball 🏀 and soccer ⚽ and I love watching weightlifting 🏋️ track & field, and wrestling 🤼.

Die-hard fan of Giannis Antetokounmpo 🇬🇷 (Miami Heat) and Kylian Mbappé 🇫🇷 (Real Madrid).

Avid reader 📚 — social psychology, behavioral psychology, and social ideology.

In my spare time I follow research on AI / large language models, computer science, HCI, and human performance.