Yichuan Zhang

IP Collaboration Candidate Screening

National popularity rankings can hide whether an IP is unusually relevant to a specific audience. I co-created a decision-support prototype that separates market potential, audience lift and structural affinity, then tests the shortlist under held-out evaluation. The result is an evidence-qualified candidate list for human due diligence, with rankings withheld when support is weak.

Co-creator — Team とあるパンダ · 9th Data Science Contest · 2026

LLM Agent–Embedded Cruise Procurement Platform

Cruise procurement work is fragmented across documents, spreadsheets, email and LINE, making handoffs and operational data difficult to trace. I designed and implemented a full-stack platform with seven bounded business domains and a tool-using LLM agent whose proposed actions require user approval. The architecture connects eight procurement stages and 30 relational entities, supporting PDF-to-RFQ and LINE agent flows without claiming unverified production gains.

Full-stack Engineer · 2024 – Present

RealityAI

Property filters capture price and bedroom count but miss preferences such as natural light, quiet streets and spatial layout. I validated floor-plan search signals and designed the CLIP retrieval, Qdrant search, rule filtering, explainable reranking, recommendation APIs and interaction architecture. The team produced an internally tested pilot for natural-language discovery and iterative preference refinement before development ended after the founder exited.

Floor-plan POC, multimodal retrieval, ranking, backend API & interaction architecture · 2025

CAS 2026 Paper Network

A flat conference programme makes related work difficult to discover across sessions and research areas. I independently designed and built the complete semantic pipeline and interactive 3D network in one month. The official CAS 2026 experience connected 51 papers through 125 semantic relationships, was used during the conference and received positive feedback.

Sole Developer · CAS 2026

Loomi — AI Creative Canvas

Most generative tools jump from a prompt to a finished image, skipping the process of collecting and recombining creative material. I co-developed a tree-based canvas concept that retrieves texture, brightness and other material attributes from a database to guide new creative directions. The concept connects a reusable material library, controlled remixing and branching iteration, keeping alternative directions and creative decisions with the user.

Prototype & Concept — Team とあるパンダ · 2026

Yujian Reading Circle

Urban social connection often remains bounded by school, workplace, age and professional identity. I founded and operated a Tokyo reading community, shaping its concept, visual identity, programming, guest coordination, hosting, communications and website. It became a recognisable recurring community, expanded from reading into cross-topic salons and helped establish the groundwork for a wider cultural community.

Founder / Community Builder / Brand Designer / Operator · 2022 – 2024

Interdisciplinary Fusion Project

Engineering, social-science, humanities and design students at UTokyo rarely share a durable space for collaboration. I led the organisation of seminars, mixers and co-creation workshops and helped build the digital platform that records member interaction and projects. The current outcome is a working collaboration format and project platform; participation scale and longer-term impact remain to be documented.

Lead Organiser · 2025 – 2026

Calibrated Intervention in Human–AI Collaborative Forecasting: How Modification Intensity and Direction Interact to Determine Performance

Human intervention does not automatically improve an AI forecast; confident corrections can amplify error when direction or magnitude is poorly calibrated. I led an eye-tracking study that operationalised correction agency, effectiveness and forecast-value-added, then analysed their interaction with mixed-effects models. The study shows that correction direction and intensity must be evaluated together: intervening more can improve or substantially worsen performance.

Zhang, Y., Hiekata, K., & Nakashima, T. · International Journal of Human-Computer Interaction · 2026 · Published

Human-AI Collaboration in ATC — Contrail Avoidance with Eye Tracking

Air-traffic controllers must consider climate-relevant contrail guidance without compromising safety or adding unmanageable workload. I am designing an eye-tracking study of when recommendations are read, deferred or overridden during representative radar-display tasks. The current output is an experimental framework for attention-aware decision support; empirical outcome evidence is still being collected.

Urban Flood Risk Simulation

Transit-access research needs a defensible estimate of where intense rainfall places flood hazard on the street network. I built a frozen 5 m hydrodynamic workflow and evaluated its spatial ranking against reported cases from two storms and independent susceptibility evidence. All eight evaluated non-favela cases ranked above their local background median, with a 94.27% mean local hazard percentile, while water-depth accuracy remains unvalidated.

Flood Simulation & Hazard Modelling · NetMob 2026 · accepted