HouseHub+

Inspiration

Public support systems, especially housing and community benefits, are often fragmented and painfully slow. People in crisis are forced to wait in long queues, while administrators drown in paperwork, struggling to verify documents and prioritize the most urgent cases fairly. We were inspired to build HouseHub+ to fix this broken pipeline and reduce friction in accessing essential services. We wanted to create an intelligent triage system that speeds up resource allocation without removing human empathy or compromising citizen privacy.

What it does

HouseHub+ is a secure, AI-powered housing benefits navigator divided into two seamless experiences:

1. The Citizen Portal: A secure interface where applicants submit their demographic details, general location, and official documents (Income/Lease). 
2. The Admin Dashboard: An actionable decision-support hub for government or NGO workers. Instead of a messy inbox, admins see a sorted queue prioritized by a Machine Learning-generated "Need Score". 

Crucially, it features a Red Alert Queue. If our computer vision models detect a forged document or fail to read it, the application isn't blindly rejected by the AI. Instead, it is routed to this queue for manual human inspection, keeping the human firmly in the loop.

How we built it

We architected HouseHub+ with a strict focus on modularity and security:

• Frontend & UI: Built using Streamlit for a highly interactive, responsive dashboard, paired with Folium to generate localized heatmaps.
• Privacy & Security Layer: Implemented cryptography (Fernet) for AES-256 encryption of names, and hashlib for SHA-256 one-way hashing of National IDs.
• Vision AI: We combined OpenCV using Error Level Analysis (ELA) to detect image manipulation/forgery, and EasyOCR (initialized for both English and Eastern Arabic numerals) to extract financial figures inclusively.
• Priority & Geo-AI: We trained a Scikit-Learn RandomForestRegressor on synthetic demographic data to output a normalized Need Score (0-100) and Tier. We also deployed a KMeans clustering model on GPS coordinates to map high-demand districts without pinpointing individual addresses.
• Database: Managed via SQLite with strict ON CONFLICT schemas to decouple personally identifiable information (PII) from assessment data.

Challenges we ran into

Building a production-ready MVP in a few days presented several hurdles:

• Memory Management: Loading heavy ML models and OCR readers concurrently caused memory leaks. We solved this by implementing strict singleton patterns and @st.cache_resource decorators.
• OCR Reliability & Over-reliance: We realized that poor-quality images caused the OCR to output zeros, which could lead to unfair scoring if the system blindly trusted the AI. We mitigated this over-reliance risk by implementing a strict AI Guardrail: if OCR extraction fails, the AI stops scoring and immediately flags the application for human review.
• Backward Incompatibility: Upgrading our encryption from basic encoding to military-grade AES-256 risked breaking the database. We built graceful fallbacks using try/except blocks to handle legacy records securely.

Accomplishments that we're proud of

1. Privacy-by-Design: We successfully decoupled citizen identities from their financial data. An admin only sees a hashed User_Token and a Need Score. The applicant's real name is only decrypted the exact moment an admin clicks "Approve".
2. True Human-in-the-Loop: We successfully mitigated the risk of automated rejections. Our AI acts only as an advisor and fraud-detector, leaving the final, empathetic decision to the human administrator.
3. Inclusivity: By ensuring our OCR natively supports Eastern Arabic numerals alongside English, we built a tool that respects local nuances and doesn't marginalize non-English speaking citizens.

What we learned

We learned that "Responsible AI" is not just a buzzword; it requires deliberate software engineering choices. We learned how to balance the speed of Machine Learning with the necessity of privacy (using hashing/encryption) and fairness (using inclusive OCR and human oversight) to provide real impact and decision value.

What's next for HouseHub+

Our immediate next steps include migrating the backend to PostgreSQL to handle millions of concurrent applications, deploying the containerized app via Docker to AWS, and upgrading our document verification from basic OCR to deep-learning-based Document AI (like LayoutLM) for full structural understanding of governmental forms.

Built With

• cryptography
• docker
• easyocr
• folium
• opencv
• pytest
• python
• scikit-learn
• sqlite
• streamlit