ArchFin AI: BIM MPT Workspace

AI workspace built on React & Revit 2027 (Nice3point). Extracts market data via NVIDIA Nemotron on Nebius, runs MPT matrix math for optimal urban zoning, and updates the BIM canvas live.

Comment

🔬 Research Foundation & Version History

ArchFin AI: BIM MPT Workspace is the direct Version 2 production evolution of my original prototype built for the All Things Agentic Hackathon.

📚 Academic Foundation

This framework is built upon robust mathematical and architectural foundations detailed in my scientific research paper: "Generative BIM Layout Optimization via Modern Portfolio Theory." While Version 1 proved the core thesis using local scripting variables, this Version 2 upgrade completely transforms the workflow into a professional, enterprise-ready decoupled software system built natively for cloud-accelerated AI infrastructure.

This project is built on research presented at ICICPE 2026:

Paper: "Generative BIM Layout Optimization via Modern Portfolio Theory"

  • Conference: 10th International Conference on Interdisciplinary Research in Computer Science, Psychology, and Education
  • Date: August 19-21, 2026
  • Venue: Chiang Mai, Thailand
  • Authors: Sherif Ahmad Magdaldin, WorldQuant University

Key Contribution: Adaptation of Markowitz Modern Portfolio Theory to spatial allocation constraints while maintaining nonnegative weights (physical feasibility).

Paper Review

💡 Inspiration

Traditional architectural design tools and real estate financial realities are completely isolated from one another. When global market shocks occur—such as lumber or construction steel prices spiking, or shifting localized asset risk indices—architects and developers have no interactive method to see how it threatens overall project layout viability. We built this workspace to bridge this gap, allowing urban planners to dynamically map out the mathematical "Efficient Frontier" of spatial allocations in real-time.

⚙️ What it Does

ArchFin AI introduces a decentralized multi-agent workflow that acts as an intelligent co-pilot for urban zoning design variations:

  • Macro-Inference Agent: Interacts with high-fidelity NVIDIA Llama-3-Nemotron models on Nebius Cloud to ingest raw market news or commodity updates and convert them into live asset risk coefficients.
  • Quantitative MPT Agent: Runs high-performance matrix inversions locally to resolve optimized allocations across Residential, Commercial, and Industrial parameters using Markowitz Modern Portfolio Theory (MPT).
  • Adversarial Inspector Agent: Acts as an automated zoning enforcement layer, continuously evaluating proposed layouts against boundary parameters and structural criteria, forcing immediate localized self-correction loops.

🛠️ How I Built It (The V2 Architecture)

To scale the original prototype into a production-grade workspace, I completely decoupled my development ecosystem into two separate, clean code nodes:

  • Frontend UI Layout (frontend/): A responsive React.js web dashboard that utilizes math.js to process vector constraints, calculate portfolio optimizations, and handle live endpoint data mapping loops.
  • Backend Core (backend/): An external application for Autodesk Revit 2027 built using the modern Nice3point framework. It initializes a non-blocking background server thread directly inside the Revit runtime lifecycle to stream results from the React frontend and instantly update parametric elements on the active canvas.

Quick Start Scenarios on Live Demo

Click one of these preset market scenarios to see real-time MPT optimization:

  1. "Steel Price Inflation" → See how industrial zone allocation shifts
  2. "Residential Boom" → Watch commercial allocation decrease
  3. "Multi-Region Shock" → Observe portfolio rebalancing across all zones

✅ Live Working Deployment

The complete system is deployed and fully operational at:

archfin-ai-bim-mpt-urban-optimize.ai.studio

Key Working Features Demonstrated:

  • NVIDIA Llama-3 Nemotron 70B integration (live inference)
  • Markowitz MPT solver computing optimal allocations in real-time
  • Interactive BIM canvas with dynamic spatial visualization
  • Multi-agent orchestration (all agents operational)
  • Real-time macroeconomic signal processing

🚧 Challenges I Ran Into

Managing intense matrix calculations alongside a heavy desktop CAD application without causing screen lag or application crashes was a significant technical hurdle. We resolved this constraint entirely by decoupling the heavy multi-agent execution loop and matrix math over to the React application thread, leaving our C# Revit layer to serve as a fast, highly reactive rendering client.

🏆 Accomplishments That I'm Proud Of

I succeeded in constructing a thread-safe connection pipe between a React browser dashboard and a native desktop Revit 2027 instance. Seeing my agents autonomously process textual macroeconomic shocks, debate spatial zone constraints, and successfully drive live architectural ribbon updates is an incredible milestone.

🚀 What's Next for ArchFin AI

I plan to introduce live, web-connected web scraping utilities to ingest financial commodity indexes automatically. I am also planning to expand my dashboard with an interactive 3D model viewport utilizing Autodesk Data Exchange APIs for absolute cloud-to-desktop parity.

  • [Q4 2026] Live commodity index integration (Bloomberg/FRED APIs)
  • [Q1 2027] Interactive 3D cloud viewer via Autodesk Data Exchange
  • [Q2 2027] Multi-region Revit synchronization (cloud-to-desktop parity)
  • [Ongoing] Academic publication in peer-reviewed journal (target: Computers in Industry, Q1 Impact Factor: 9.2)

Built With

  • 8autodesk
  • ai
  • apimath.jsnvidia
  • llama-3-nemotronnebius
  • nebius
  • nvidia
  • react.jsc#.net
  • revit
Share this project:

Updates

Submission history