Originally written as my submission to the OpenAI Open Model Hackathon. The project extends AskStreets; the code lives on GitHub.
DreamStreets extends AskStreets with GPT-OSS-120b’s advanced reasoning. It uses OpenAI’s open-weight 120B model to perform chain-of-thought reasoning over geospatial questions, generating and executing graph algorithms and spatial queries to provide actionable insights for both urban planning and humanitarian contexts.
What It Does
DreamStreets leverages GPT-OSS-120b to:
- Generate and execute NetworkX graph algorithms for network analysis
- Create spatial SQL queries for POI and facility analysis
- Perform chain-of-thought reasoning to break down complex urban planning questions
- Provide actionable insights for urban planning and humanitarian response
Key Applications
- Urban Planning: optimal business locations, infrastructure vulnerability assessment
- Public Health: emergency response gap identification, route planning
- Humanitarian Response: refugee camp resource placement, flood preparedness
- Emergency Management: evacuation planning, critical bottleneck identification
Technical Architecture
- LLM: GPT-OSS-120b (120B parameters) via Ollama, completely offline
- Scientific Computing: RAPIDS AI 25.10a with CUDA 12.9 for GPU-accelerated data processing
- Graph Analysis: NetworkX with OSMnx for geospatial enhancement and nx-cugraph for GPU acceleration
- Spatial Database: DuckDB with spatial extensions
- Agent Framework: LangGraph ReAct agent with specialized tools
- Data Source: OpenStreetMap via OSMnx
Why GPT-OSS-120b
This project explores OpenAI’s most powerful open-weight model for geospatial reasoning. Key advantages:
- Complete offline operation, critical for field deployment in low-connectivity areas and under data regulation
- Transparent chain-of-thought, so we can audit the full reasoning process for debugging and trust
- Proven tool use on agentic workflows
- Efficient MoE architecture: only 5.1B active parameters per token despite 117B total
The model’s unsupervised chain-of-thought is particularly valuable for urban planning: decision-makers can audit the reasoning behind infrastructure recommendations. Running locally ensures data sovereignty for sensitive government and humanitarian datasets.
Why RAPIDS + LLMs
This project demonstrates the synergy between established scientific computing tools and language models. The RAPIDS ecosystem provides battle-tested GPU acceleration for data manipulation, while GPT-OSS-120b adds natural language understanding and code generation. The combination lets domain experts leverage complex algorithms without writing code, integrating smoothly with their GIS workflow.
Deployment
Deployed on Runpod with the rapidsai/notebooks:25.10a-cuda12.9 base image and a JupyterLab environment. Hardware: NVIDIA A100 SXM (80GB VRAM), 250GB system RAM, at roughly $1.55-1.74/hr. This was my first experience running large models on enterprise GPUs.
Key Achievements
- Chain-of-thought reasoning for complex spatial problems
- Dynamic code generation for NetworkX algorithms and SQL queries
- Multi-tool orchestration combining graph and database analysis
- Complete offline operation, critical for field deployment and compliance
On performance: it analyzes networks with hundreds of nodes in seconds, initializes the model once per session (30-second warmup), and processes complex multi-step queries with up to 25 reasoning iterations, with no ongoing API costs after setup.