RentSentry

AI-powered rental fraud detector - paste any Craigslist URL or listing text, or use the browser extension on Facebook Marketplace, and get an instant trust score, red flags, and neighborhood insights.

Comment

Inspiration

Rental scams cost Americans hundreds of millions of dollars every year. Victims are often first-time renters, students, or people relocating to a new city — people who can least afford to lose a security deposit or month's rent to a fraudster. We wanted to build something that anyone could use before wiring money to a stranger on the internet.

What We Built

RentSentry is a rental listing fraud detector. You paste a Craigslist URL (or raw listing text) and get back:

  • A trust score (0–100) — a weighted composite of LLM suspicion analysis and price heuristics
  • Red flags — short, plain-English fraud signals extracted by the LLM (e.g. "Wire transfer required", "Landlord claims to be overseas")
  • Accessibility signals — positive location signals like nearby transit, parking, and ADA features
  • A clean listing preview — the LLM rewrites cluttered ALL-CAPS, emoji-heavy Craigslist text into professional prose
  • Reverse image search links — Google Lens and TinEye links under every listing photo so you can spot stock images in one click
  • A browser extension that injects an inline trust badge directly on Craigslist and Facebook Marketplace pages

The trust score formula is: $$\text{trust_score} = 100 - ((\text{llm_score} \times 0.6) + (\text{price_score} \times 0.4))$$

How We Built It

The backend is a FastAPI service with three modules:

  1. scraper.py — fetches and parses Craigslist listings using httpx + BeautifulSoup
  2. llm_analysis.py — sends listing data to OpenAI gpt-4o-mini (with Anthropic claude-haiku-4-5 as a fallback) and returns a suspicion score, red flags, and a paraphrased description
  3. main.py — orchestrates everything, applies price heuristics, and returns a unified JSON response

The frontend is intentionally lightweight — a single index.html with vanilla JS, no build step needed. It includes a demo mode that works entirely in-browser without the backend.

We also built a Firefox browser extension (Manifest V3) that uses a MutationObserver to detect listing pages on Craigslist and Facebook Marketplace and inject the RentSentry panel automatically.

The service is deployed on Render.

Challenges

  • Craigslist anti-scraping: Craigslist aggressively blocks automated requests. We had to carefully mimic browser headers and handle edge cases like missing price fields, QR code footers in descriptions, and lazy-loaded images.
  • LLM reliability: LLMs occasionally return malformed JSON or refuse to score borderline listings. We wrapped every LLM call in a safe fallback that returns a neutral default dict so the app never crashes.
  • Facebook Marketplace scraping: Facebook uses a React-rendered DOM with no stable IDs. We had to use a MutationObserver in the extension's content script to wait for the listing content to appear before extracting it.
  • Price heuristics without a real API: We built a Boston-specific median rent table by hand to power the market price score. Scaling this to other cities is on our roadmap.

What We Learned

  • How to structure a multi-model LLM pipeline with graceful fallbacks
  • How to write Manifest V3 browser extensions that work across dynamically rendered pages
  • How to balance speed vs. accuracy when chaining async scraping + LLM calls

Built With

  • anthropic-api-(claude-haiku-4-5)
  • beautifulsoup4
  • fastapi
  • firefox-extension-(manifest-v3)
  • html/css
  • httpx
  • javascript
  • openai-api-(gpt-4o-mini)
  • python
  • render
Share this project:

Updates