UltimateDESIGN
AI Urban Design & Micro-Renewal Multi-Agent Decision Platform
An AI urban design and micro-renewal decision platform empirically grounded on a 170.2 ha Changchun site. Converges 17 design stages into 3 core workbenches with GIS-to-ControlNet spatial alignment and 3-stakeholder LLM negotiation.
- Python 3.12
- Streamlit
- GeoJSON / GIS
- ControlNet
- DeepSeek
- Gemma Vision
- BM25 + Vector RAG
- Pytest
- Plate
- PLATE 02
- Role
- Primary Author
- Status
- Empirical Build · 356 Pytests Passing
- Year
- 2026
- Scope
- Primary Author: Independently engineered the 17-stage spatial logic framework, GIS-to-ControlNet alignment pipeline, 3-stakeholder LLM simulation, and 356-test Pytest verification suite.
Interactive HTML Slide Deck
Navigate with controls, keyboard arrows, slide indicators, or Fullscreen presentation mode.
Renders the 170.2 ha empirical renewal site in cinema-dark wireframe precision with parcel boundaries and street vectors.
- 170.2 ha empirical GeoJSON parcels and footprints
- Diagnostic metrics anchored to real spatial coordinates
- 1:500 architectural drafting clarity in cinema-dark mode
- 17 Logical Design Stages
- 170.2 ha Changchun Empirical Site
- 26 Standard A3 Plate Templates
- 356 Automated Pytest Checks
Context & Engineering Constraints
Unconstrained generative image tools ignore real GIS geometry and urban codes, causing road-network drift, invalid land-use metrics, and stakeholder misalignment unusable in professional planning reviews.
System Architecture
- 01
17 Logical Stages × 3 Core Workbenches
Consolidates stages 00–16 into three integrated Streamlit workbenches: Spatial Diagnosis Sandbox, Multi-Agent & AIGC Studio, and Digital Regulatory Delivery.
- 02
GIS → Vector → Raster → ControlNet Alignment
Extracts exact GeoJSON parcels and street alignments into deterministic control maps, pairing 6 renewal strategies with ControlNet to eliminate spatial drift.
- 03
3-Stakeholder LLM Game & Dual-Critic Loop
Simulates Residents, Developers, and Planners to reach balanced consensus, audited by Gemma Vision visual critique, DeepSeek regulatory RAG, and 356 Pytest checks.
Engineering Highlights
- Validated end-to-end on a 170.2 ha Changchun urban renewal district from GIS spatial diagnosis through 26 national-standard A3 plate templates.
- Locks parcel footprints and road topology from GIS vectors across 6 micro-renewal strategies to prevent geometric hallucinations in AIGC renders.
- Models multi-turn Resident, Developer, and Planner negotiation as a transparent consensus radar rather than opaque single-prompt generation.
- Pairs a Gemma Vision + DeepSeek dual-critic review loop with 356 passing Pytest checks to safeguard spatial data integrity.