Return to Home

Feiyang Li

Shenzhen, China·384304995lfy@gmail.com·www.feiyanglee.com

Research Interests

Artificial general intelligence, in particular agentic and generative systems that create and reason about 3D spaces (buildings, urban districts, game levels) while staying controllable, constraint-aware, and reproducible.

Education

Columbia University, Graduate School of Architecture, Planning and Preservation May 2024 – May 2025
M.S. in Advanced Architectural Design · New York, USA
Wuhan University Sep 2018 – Jun 2023
B.Arch. in Architecture · Wuhan, China · GPA 3.7/4.0 · Outstanding Graduate (top 10% of the graduating class)

Research Experience

Procedural Generation of High-Density Open-World Urban Districts 2026
System Designer, Tencent
  • Build a procedural generator for a high-density walled-city district in an open-world game, with rule-based layouts and full interiors in which every room is reachable through corridors, so the result is a traversable town.
  • Define which rooms make up each home, and use an LLM agent to generate a structured JSON description of each home and its occupant. A second agent operates Blender to produce the art assets.
  • Generate irregularly shaped lots and building footprints instead of rectangular grids, and plan relations between neighboring buildings (apart, fused, bridged) before geometry.
  • Verify every output with an independent validator, report infeasible placements as unresolved instead of patching them, and seed all random streams so results are reproducible.
Multi-Agent 3D Level Generator 2026
System Designer, Tencent
  • Designed the end-to-end workflow of a multi-agent LLM level generator: five stages from topology to detail, each producing a structured description that designers can review and use directly, coordinated by a lead agent that delegates to specialist agents.
  • Defined a verification workflow with acceptance criteria at every stage, including a simulated playthrough whose key-node screenshots a vision-language model checks for spatial plausibility and guidance clarity. Failures are sent back to the earlier stage instead of accumulating downstream.
  • Built a browser-based block-out sandbox that exports to Unreal Engine 5, and used the system to produce linear level block-outs. Scored output with a design rubric against two baselines, single-prompt LLM generation and hand-built levels, and found it markedly better than the former.
Walkability Analysis of the Manhattan Street Network 2024
Course Project, Computational Modeling, Columbia GSAPP
  • Built a street-level walkability analysis covering all of Manhattan, integrating NYC Open Data on pedestrian and bicycle networks, building heights, and street trees into one dataset with Grasshopper and QGIS.
  • Computed shade from buildings and trees through shadow analysis, and measured greenery from tree data, for every street segment.
  • Combined these factors into a weighted walkability score for each street and visualized the results in a video and a route map recommending where to walk or cycle.
Multi-Objective Parametric Design Optimization 2023
Undergraduate Thesis, Wuhan University · Advisor: Xuecheng Gui
  • Formulated building massing as a three-objective optimization problem (landscape view, daylighting, floor area ratio) and solved it with a multi-objective genetic algorithm in Rhino/Grasshopper, keeping the objectives separate to obtain a Pareto front of trade-offs rather than a single weighted score.
  • Selected the final scheme from the Pareto front, improving landscape view quality by about 17% over the manual baseline while maintaining daylighting and floor area ratio.
  • Ran a second-stage genetic search driven by Ladybug/Honeybee simulation, reducing simulated energy use by 6%.

Professional Experience

Tencent — Level Designer Apr 2026 – Present
Shenzhen, China
  • Use LLMs daily for level concepts, design documents, and configuration data, working with programmers and artists on production.
  • Defined quantitative design standards for an open-world driving game and built Python and Unreal Engine 5 tooling that generates test scenes to check them against playtest evidence.
Tianhua Architecture Design Institute — Architectural Designer Oct 2025 – Mar 2026
Shanghai, China
  • Designed residential schemes and developed an AI-assisted rendering workflow for design iteration and presentation production.

Technical Skills

Honors

Leadership

Committee Member, Game Development Club, Columbia University Jun 2024 – May 2025
  • Co-organized LionJam, a campus game jam for more than 100 participants, with speaker sessions and a partnership with Tencent.
President, Chunhui Public Welfare Center, Wuhan University 2019 – 2022
President from 2021
  • Led rural volunteer-teaching programs with more than 100 volunteers, coordinating students, schools, and local partners. The center was named one of Wuhan University's Top Ten Student Organizations, and its work was reported by Xinhua News Agency.
Department Head, Student Innovation and Entrepreneurship Practice Center, Wuhan University Jun 2020 – Jun 2021
  • Ran promotion and operations for student innovation competitions, including Internet Plus, the Shenzhen Innovation and Entrepreneurship Competition, and a 3D-printing design competition.