Prospective PhD student, Human–AI Interaction · Founder of Usercase · Seoul
I build AI agents that study people: they watch what people do, ask why, and notice when words and actions disagree.
I trained in civil and mechanical engineering at Sungkyunkwan University, graduating summa cum laude in 2022, and built AR and computer-vision tools in its structural health monitoring lab. Since 2025 I have run Usercase, where an AI interviewer asks people about what they just did inside a product. In a PhD I want to study how AI agents can learn about people faithfully, and when a simulated person can stand in for a real one.
How can AI agents study people faithfully: asking from what people actually do, and knowing when a simulated person can stand in for a real one?
Follow-up questions built from what a person just did, not only from what they report; when to step in, and how to ask without leading.
Evidence: Usercase
Agents built from a person's own interview; measuring where they answer like the real person and where they drift.
Evidence: Dalm · willu
Comparing design with as-built structures through vision, 3D scans and AR: measuring the gap between a model and reality, and showing its error honestly.
Evidence: HoloLens · crack detection
KeywordsHuman–AI Interaction · AI-moderated interviews · Generative agents · Simulations of people · Behavioral data vs. self-report · Human-in-the-loop systems · Spatial computing · Civil × Mechanical × AI
01 · AI interviewers grounded in behavior
An AI user-research platform that opens an interview inside the product when behavior shows friction. In a 50-user live study, 48 reached the goal by different paths, and interviews to report took one day instead of about two weeks. People often reported difficulty on tasks they had completed, so the AI moderator was rebuilt to ask from behavior.
Interactive re-creation of a session ▶02 · Simulating people with agents
Stanford's generative-agent interviews (Park et al., 2024), rebuilt in Korean: an AI interviewer runs a life-story interview of up to 71 questions, then builds an agent that answers like the participant.
Learned: Long interviews stayed on protocol only when the script lived in code and the model decided just the follow-ups.
Open Dalm ↗02 · Simulating people with agents
A voice interview builds your AI twin; twins go on simulated dates with other twins to suggest who you should actually meet.
Learned: A negative result: when one model played both people, simulated dates could not separate pairs we expected to work from pairs we expected to fail (mean 55% vs. 49%; 10 demo-persona pairs, 200 scenes). That pushed the design toward an independent agent per person.
Open willu ↗03 · Where it started
Scanned full-scale precast slabs, compared them with the design model, and overlaid the target in the worker's view with Unity and marker registration. The overlay worked; tracking drift ruled out precision placement.
Interactive re-creation ▶03 · Where it started
MATLAB pipelines tuned across lighting and surface conditions rather than for one good image; depth checks on 3D scans; and a computer model driving an Arduino stepper rig through LabVIEW.
Interactive re-creation ▶Sangwan (Richard) Kim · Seoul · sangwan.kim
