GEOGRAPHY · TECHNOLOGY · PEOPLE

Follow curiosity.
Understand the world.
Build what matters.

I think of myself as a curious generalist. Geography is where I begin to observe the world; technology is how I understand and change problems; people are the measure I use to decide what is worth building.

See what I’ve built See my experience

Move to exploreWuhan · Tokyo · Hangzhou · Beijing

Get to know me

00 / ABOUT

Between the human
and the technical.

I studied Geospatial Information Engineering in both my bachelor’s and master’s programs. Geography taught me to pay attention to people, cities, space, and relationships; programming lets me turn judgment into tools, systems, and testable prototypes.

From WebGIS and spatial analysis to NL2SQL and Agents, I conduct research, write code, run services, and shape products. I do not want technology detached from real people, nor do I want to stop at discussing ideas. Curiosity opens the question, a human perspective sets the direction, and technical ability lets me make the idea real.

01 / EXPERIENCE

Paths I’ve taken

Across these three experiences, I conducted research, wrote code, and worked with real users to move products forward.

May 2026 — PresentBeijing

New Oriental Education & Technology Group

Internet Center

AI Product R&D Intern

Scaled the production of interactive card-based courses by turning teachers’ methods into reusable Skills, generating editable drafts with an Agent, and iterating through card-level evaluation.

Built a course-building Agent with Pi Agent SDK, covering requirement clarification, course planning, card generation, and draft creation on the live authoring platform.
Analyzed interviews, historical course-building records, and teachers’ prompts to turn card rules, selection logic, and content requirements into subject-specific Skills.
Designed a card-level LLM-as-Judge evaluation process and captured teacher edits as failure cases; shifted work from manual assembly to review and editing, cutting production time by about 50%.

Feb 2026 — Apr 2026Hangzhou

Zhejiang Jingzhunxue Technology Co., Ltd.

Product Department

AI Product Manager Intern

Built an iteration process for an AI tutoring device, spanning evaluation criteria, bad-case diagnosis, prompt improvement, and automated testing.

Analyzed more than 50 expert review records and tutoring guidelines, breaking quality into guidance, feedback, pacing, summaries, and emotional support with structured labels.
Used simulated students at different ability levels and real questions to diagnose conflicting rules and examples, helping raise the pass rate for primary-school math to 92%.
Integrated with the prompt platform API and built dialogue-simulation and automated-evaluation tools, reducing one evaluation cycle from roughly two hours to ten minutes.

Sep 2023 — Dec 2025Tokyo

LocationMind Inc.

R&D Division

AI Product & R&D Intern

Started with natural-language access to geospatial databases, expanded into geospatial analysis Agents, and eventually brought those capabilities into a workplace-facing digital employee prototype.

Proposed dynamic schema mapping and human-in-the-loop clarification; Monkuu achieved 56.2% accuracy on KaggleDBQA, 13.8 points above the previous best, and was published in IJGIS as a first-author paper.
Built a geospatial analysis Agent with PocketFlow and ReAct, exposing professional tools through QGIS-MCP and reaching a 73.3% prototype evaluation pass rate.
Developed the GeoClaw digital employee prototype and improved accuracy from 73% to 95% across 109 real-world tasks.

EDUCATION

Education

M.Eng. CandidateSep 2024 — Jun 2027

China University of Geosciences (Wuhan)

Geospatial Information Engineering · Project 211 University

Top 5% in the program; recipient of the university First-Class Scholarship.

Visiting Research Program, Center for Spatial Information Science (CSIS), The University of Tokyo ↗Visiting-program advisor: Prof. Ryosuke Shibasaki ↗

B.Eng.Sep 2020 — Jun 2024

China University of Geosciences (Wuhan)

Geospatial Information Engineering · Project 211 University

Top 5% in the program and admitted to graduate study by recommendation; participated in a national undergraduate innovation program and received a software copyright and the CUG Talent Scholarship.

02 / SELECTED WORK

Things I’ve built

Each project began with a geographic problem: understand the data, call the right tools, and bring the result into a real workflow.

01NL2SQL · GeoAI · First-author paper

Monkuu

A natural-language interface for people who do not write SQL. Dynamic schema mapping handles large schemas, while human-in-the-loop clarification resolves ambiguous spatial intent.

56.2% on KaggleDBQA, 13.8 percentage points above the previous best result.

Read paper ↗

02ReAct · QGIS-MCP · Complex spatial tasks

Geospatial Analysis Agent

An Agent that selects spatial analysis tools, recovers from execution errors, and returns maps, charts, and analytical results to the user.

73.3% prototype evaluation pass rate, establishing the capability base for later digital employee work.

03Agent · Slack Bot · CLI

GeoClaw Digital Employee

A workplace interface that brings geospatial analysis out of the command line, allowing users to submit tasks and receive results through conversation.

Accuracy improved from 73% to 95% across 109 real-world tasks.

03 / BUILT FOR MYSELF

Built because
I needed them

Tools, methods, and infrastructure I maintain outside my primary work.

Open-source tool · 2026

cc-daily-report

Collects incremental Claude Code conversations and turns them into a daily review and Morning Review, preserving context from intensive human–AI collaboration.

View project ↗02

Method & Skill · 2026

ADD

Reframes multi-agent coding as a management problem built around isolated workspaces, document-based communication, review gates, and structured delivery.

GitHub ↗03

Maintained since 2023

UrbanComp Digital Infrastructure

Built and maintain more than eight services for the lab, including AI Chat, its website, collaborative documents, private storage, computing services, and a forum.

Visit UrbanComp ↗04

Personal experiment · 2026

Hot 100 Memory Plan

Turns algorithm practice into spaced review, scheduling problems by whether I can solve them independently, with hesitation, or not yet.

Try it online ↗

04 / RESEARCH

Research is another way
to understand the world.

My work extends from human–database interaction for geospatial data to multimodal land-use classification, geographic representation learning, and global risk analysis.

ResearchGate ↗

2025First author