> For the complete documentation index, see [llms.txt](https://propblock-ai.gitbook.io/propblock-ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://propblock-ai.gitbook.io/propblock-ai/rwa-protocol-overview/ai-agent-operational-architecture.md).

# AI Agent Operational Architecture

PropBlock’s AI Agent operates based on a **three-layer architecture**, comprising the Data Layer, Analytics Layer, and Application Layer. These layers are responsible for data collection, data analysis, and project execution, respectively.\
\
***The operational structure of the AI Agent is as follows:***

### 1. Core Functional Module Expansion

#### **1.1 Data Layer**

The Data Layer serves as the AI Agent’s “sensory system”, responsible for collecting and structuring real-world data from multiple sources, providing the fuel for upper-layer analysis.

**Data Source Integration**

<table><thead><tr><th width="161">Submodule</th><th width="245">Details</th><th>Technology/Tools</th></tr></thead><tbody><tr><td>Market Data</td><td>Real-time API access to global real estate prices, updated monthly; rental data is scraped from platforms like Airbnb, with deduplication applied.</td><td>Scrapy, Apify, Restful API</td></tr><tr><td>Geographical Data</td><td>Integration of real-time traffic data from Google Maps API; POI data categorized (Education/Healthcare) and analyzed within a 500m buffer zone.</td><td>Google Maps API, GeoPandas</td></tr><tr><td>Policy Data</td><td>Dynamic monitoring of government websites; OCR-based parsing of PDF policy documents, extracting key terms (e.g., “purchase restriction,” “FAR adjustment”).</td><td>Tesseract OCR, BERT Keyword Extraction</td></tr><tr><td>Economic Indicators</td><td>Connection to local National Statistics Bureau API for quarterly GDP data; integration with third-party data platforms (e.g., Wind) for employment rate indicators via Snowflake.</td><td>Snowflake, Tableau</td></tr></tbody></table>

**Data Preprocessing**

<table><thead><tr><th>Submodule</th><th width="249">Details</th><th>Technology/Tools</th></tr></thead><tbody><tr><td>Data Cleaning</td><td>Housing price anomaly detection: Identifying data deviating from market trends by more than three standard deviations using Isolation Forest.</td><td>PyOD, Scikit-learn</td></tr><tr><td>Standardization</td><td>Geocoding unified to WGS84 coordinate system; housing price units converted to “USD/m² per month.”</td><td>GDAL, Python RegEx</td></tr><tr><td>Knowledge Graph</td><td>Constructs “Land Parcel - Policy - Economic Indicator” triplets, e.g., &#x3C;Land A, Affected by Policy, Purchase Restriction 2023>.</td><td>Neo4j, Apache Jena</td></tr></tbody></table>

#### **1.2 Analytics Layer**

The Analytics Layer functions as the “brain” of the AI Agent, leveraging machine learning and operations research models to transform data into actionable insights. It upgrades traditional experience-driven decision-making into a data-driven dynamic optimization process, balancing returns and risks.

**Market Dynamics Modeling**

<table><thead><tr><th>Submodule</th><th width="249">Details</th><th>Technology/Tools</th></tr></thead><tbody><tr><td>Time Series Forecasting</td><td>Prophet model incorporates custom seasonal factors (e.g., policy cycles); LSTM utilizes an Attention mechanism to capture long- and short-term dependencies.</td><td>Facebook Prophet, TensorFlow + Keras</td></tr><tr><td>Clustering Analysis</td><td>K-means cluster optimization based on silhouette coefficient; features include “Price/Rent Ratio,” “Metro Station Density,” “School District Rating.”</td><td>Scikit-learn, Yellowbrick</td></tr></tbody></table>

**Location Decision Engine**

<table><thead><tr><th>Submodule</th><th width="249">Details</th><th>Technology/Tools</th></tr></thead><tbody><tr><td>Multi-Objective Optimization</td><td>Defines objective functions: Maximize (ROI), Minimize (Policy Risk), Maximize (Transport Accessibility); constraints include “Budget ≤ USD 100M.”</td><td>Pyomo, NSGA-II Algorithm</td></tr><tr><td>Spatial Weight Analysis</td><td>ArcGIS Network Analyst calculates commuting time from land parcel to CBD, with weighted allocation: Transport (40%), Schools (30%), Commercial Amenities (30%).</td><td>ArcGIS Pro, Python ArcPy</td></tr></tbody></table>

**Risk Assessment Models**

<table><thead><tr><th>Submodule</th><th width="249">Details</th><th>Technology/Tools</th></tr></thead><tbody><tr><td>Policy Sensitivity Analysis</td><td>Monte Carlo simulation of policy variables (e.g., interest rate increase of 0.5%–2%), outputting probability density distribution of returns.</td><td>NumPy, Matplotlib</td></tr><tr><td>Market Stress Testing</td><td>Extreme scenario settings: GDP growth decreases by 3%, unemployment rate rises by 5%, testing probability of cash flow breakdown in asset portfolios.</td><td>StressTesting Library, Custom Python Simulator</td></tr></tbody></table>

#### **1.3 Application Layer**

The Application Layer serves as the AI Agent’s “interaction interface,” translating complex analytical results into user-friendly solutions. It hides technological complexity behind intuitive interfaces, ensuring a seamless Analyze-Decision-Execution loop.

**Interactive Interface**

<table><thead><tr><th width="143">Submodule</th><th width="249">Details</th><th>Technology/Tools</th></tr></thead><tbody><tr><td>Natural Language Query</td><td>User input parsing: SpaCy for entity recognition + Rasa for dialogue management, supporting multi-turn interactions (e.g., “Filter out old properties before 2020”).</td><td>Rasa Framework, SpaCy NLP Library</td></tr><tr><td>Visualization Dashboard</td><td>Power BI integrates Folium map plugin, supporting interactive heatmaps (colored based on ROI/risk level).</td><td>Power BI Embedded, Folium</td></tr></tbody></table>

**Automated Decision Support**

<table><thead><tr><th width="180">Submodule</th><th width="249">Details</th><th>Technology/Tools</th></tr></thead><tbody><tr><td>Intelligent Recommendation System</td><td>Item-based collaborative filtering, similarity calculation: Euclidean Distance (Price Trends) + Jaccard Index (Policy Tags).</td><td>Surprise, Redis</td></tr><tr><td>Compliance Checking</td><td>Drools rule engine verifies “Land Use Type matches Planning,” e.g., “Commercial land plot must have an FAR ≥ 2.0.”</td><td>Drools, MySQL</td></tr></tbody></table>

### 2.Technical Architecture Expansion

The AI Agent enables end-to-end automation from data collection to decision support, ensuring technical robustness and user experience friendliness for complex business scenarios.&#x20;

A well-structured technical architecture is essential to support real-world project execution.

#### **2.1 Infrastructure**

<table><thead><tr><th width="164">Component</th><th width="198">Details</th><th>Reason for Selection</th></tr></thead><tbody><tr><td>Data Lake</td><td>AWS S3 bucket storage: <code>raw_data</code> (raw scraped data), <code>processed_data</code> (cleaned data in Parquet format).</td><td>High scalability; supports spatial data partitioning (by city/date).</td></tr><tr><td>Compute Engine</td><td>Databricks cluster configuration: Driver Node (16-core, 64GB RAM for GIS analysis); Worker Nodes (auto-scaling based on workload).</td><td>Integrated with MLflow for model tracking; Delta Lake ensures ACID transactions.</td></tr><tr><td>Model Deployment</td><td><p><strong>Kubernetes deployment strategy:</strong> - Predictive models (Stateless Pods, HPA auto-scaling). </p><p></p><p>- GIS Services (StatefulSet fixed nodes).</p></td><td>High availability; Istio enables model canary releases.</td></tr></tbody></table>

#### 2.2 Tech Stack

| Domain              | Tech Stack                                                                                                                       | Key Dependencies                                                                              |
| ------------------- | -------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------- |
| Geospatial Analysis | <ul><li>Spatial DB: PostGIS (supports GIS topology queries). </li><li>Dev Tools: QGIS Plugin (Python + QT Framework).</li></ul>  | PostGIS spatial indexing accelerates queries; QGIS plugin enables custom land scoring tools.  |
| Machine Learning    | <ul><li>ML Pipeline: Scikit-learn Pipeline. </li><li> Hyperparameter Tuning: Optuna + Dask distributed search.</li></ul>         | Optuna’s TPE algorithm improves efficiency by 50%+ over Grid Search.                          |
| Business Logic      | <p></p><ul><li>Web Framework: Django REST Framework. </li><li>Asynchronous Tasks: Celery + RabbitMQ (Priority Queues).</li></ul> | Celery workers separate urgent tasks (e.g., real-time risk assessment) from batch processing. |

#### 2.3 Security & Compliance

| Requirement    | Implementation                                                                                                                                                       | Technical Details                                                                                          |
| -------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- |
| Data Masking   | Homomorphic Encryption (Microsoft SEAL) protects user phone numbers/IDs while enabling encrypted computation on aggregate indicators (e.g., average age in an area). | Encryption Algorithm: CKKS Scheme (supports floating-point operations).                                    |
| Audit Tracking | Hyperledger Fabric blockchain records key operations (e.g., model parameter changes), ensuring immutable audit logs.                                                 | Chaincode defines data notarization rules; nodes include developers, regulators, and third-party auditors. |