The Compliance Layer for Real Estate AI Agent.

FairProp Inspector is a high-performance, latency-critical inference engine designed to detect Fair Housing Act (FHA) violations in real-time. Unlike legacy regex-based solutions, FairProp leverages Small Language Models (SLMs) fine-tuned on compliance datasets to understand context, nuance, and intent.

Built for the On-Device AI era, it runs efficiently on edge hardware while maintaining privacy-first architecture.

graph TD
    A[FHA Rules & Heuristics] --> B[Synthetic Generator <i>(GPT-4o Distillation)</i>]
    B --> C[(Synthetic Dataset)]
    C --> D[ModernBERT Fine-tuning <i>(BF16 / FlashAttention)</i>]
    D --> E{Model Serialization}
    E --> F[PyTorch Checkpoint]
    E --> G[ONNX Export <i>(Quantized)</i>]
    G --> H[Edge Inference <i>(Browser/Embedded)</i>]
    F --> I[Compliance API/Platform]

Part of the FairProp AI Platform ecosystem.

• SOTA Architecture: Powered by ModernBERT, delivering 8192 context length and Flash Attention backend.
• Edge-Native: Optimized for ONNX Runtime export, enabling sub-20ms latency on CPU.
• Data Engine: Includes a synthetic data generation pipeline (scripts/generate_synthetic.py) utilizing LLM distillation (GPT-4o) to bootstrap compliance supervision.
• Privacy-First: No data leaves your infrastructure. Full compliance checks happen locally.

git clone https://github.com/ZheWang-stack/FairProp-Inspector.git
cd FairProp-Inspector
pip install -e .

pip install git+https://github.com/ZheWang-stack/FairProp-Inspector.git

FairProp Inspector bridges the gap between simple regex rules and expensive cloud APIs:

Benchmarks run on Intel i7-12700K CPU with ONNX Runtime optimization.

Get started in 30 seconds:

from src.inference.predict import predict

# Detect FHA violations instantly
text = "No kids under 12 allowed"
label, confidence = predict(text, "artifacts/model")

print(f"{label}: {confidence:.1%}")
# Output: NON_COMPLIANT: 99.8%

Try it now:

python examples/quickstart.py

Explore our ready-to-run examples:

• Quick Start - 5 lines of code to get started
• Edge Inference - Production-ready with error handling and batch processing
• Batch Processing - Efficiently process multiple property listings
• Jupyter Tutorial - Interactive notebook with visualizations and performance analysis

See examples/README.md for detailed usage instructions.

Our pipeline moves away from "black box" APIs to measurable, controllable local inference.

1. Synthetic Distillation: We use gpt-4o to generate "Edge Case" violations (e.g., subtle steering like "Perfect for active adults").
2. Training: We fine-tune ModernBERT-base using bf16 precision and gradient checkpointing.
3. Inference: The model classifies text segments as COMPLIANT vs NON_COMPLIANT with probability calibration.

Train the inspector on your proprietary or synthetic data.

# Uses Flash Attention & BF16 automatically if specific hardware is detected
python src/trainer/train.py --data data/processed/synthetic.json --epochs 5 --batch_size 16

Bootstrap your dataset using our chain-of-thought distillation script.

export OPENAI_API_KEY="sk-..."
python src/generator/generate_data.py --count 1000 --output data/processed/synthetic_train.json

python src/inference/predict.py "No kids under 12 allowed in the specialized quiet zone."
# Output: [NON_COMPLIANT] 98.4% Confidence

• Training Guide - Complete guide to training custom models with GPT-4 prompt templates
• Benchmarks - Performance comparison and accuracy testing
• Examples - Ready-to-run code samples
• ROADMAP - Project development plan and quarterly goals
• CHANGELOG - Version history and release notes

We welcome contributions from the community! Please see:

• CONTRIBUTING.md - Contribution guidelines and code standards
• CODE_OF_CONDUCT.md - Community guidelines
• Issue Templates - Bug reports and feature requests

This project is licensed under the MIT License.

Built with ❤️ for Fair Housing Compliance
⭐ Star us on GitHub | 🐛 Report Bug | 💡 Request Feature