PakLand
I. Charge
Pakistani real estate platforms have a trust problem. Buyers encounter fake listings, outdated prices, and sellers who have already sold the property. Manual moderation does not keep up.
II. Investigation
PakLand is a cross-platform Flutter application that treats data accuracy as an engineering constraint rather than a moderation problem. The system enforces listing freshness with an automatic expiry rule, runs AI-based content scanning through an n8n workflow, and calculates a Trust Score for each seller. Search ranking is tied to the Trust Score, which means low-trust sellers cannot game visibility through posting volume.
III. Exhibits - Architecture
- Every listing expires automatically after 30 days. Sellers must re-verify to keep it active. This forces the catalogue to reflect the current state of available properties rather than accumulating stale entries.
- An n8n moderation pipeline runs after each listing submission. GPT-4o-Mini analyzes the listing text for fraud patterns. High-risk submissions are sent to a human review queue.
- All moderation decisions, including AI classifications and human overrides, are written to a Firestore audit log. This creates a record that can be reviewed if a seller disputes a rejection.
- Each seller has a Trust Score from 0 to 100 based on listing accuracy history, verification status, and moderation outcomes. Sellers must stay above 75 to purchase featured placement.
- Trust Score affects search ranking directly. Low-trust sellers appear lower in results regardless of listing recency. This prevents gaming the system through volume.
- The application integrates Google Maps SDK for geo-search and property location pinning.
- Infrastructure uses Firebase Authentication, Firestore, and Supabase Storage.
IV. Cross-Examination
Q: Why route high-risk AI classifications to humans rather than auto-rejecting them?
A: Auto-rejection produces false positives. A ruined building photographed for renovation might trip an image classifier. A seller in a rural area might write a description that looks unusual to the model but is accurate. Auto-rejection silently removes legitimate listings. Human review catches classifier errors before they damage seller trust.
Q: Why tie Trust Score to search ranking instead of just showing a badge?
A: Badges are visible, but they do not change behavior when the seller can still reach buyers through search. Attaching verification to search position creates a direct business incentive. Sellers who want visibility have to complete the verification process.
V. Failure Modes
- LLM rate limits halt the moderation pipeline. Listings queue until the endpoint recovers. Unreviewed listings are not published automatically.
- Google Vision false positives flag legitimate property photos. Human reviewers override incorrect classifications.
VI. Lessons Learned
- A trust badge displayed next to an unverified listing does not change user behavior. Connecting verification to search rank forces compliance because the cost of ignoring it becomes visible.
VII. What I'd Build Next
- Integrate JazzCash and EasyPaisa to handle featured listing payments without requiring international payment methods.
Verdict
ALLOW - Completed as an academic project. Authentication, AI moderation pipeline, Trust Score, geo-search, and property booking are all implemented.