RentSentry

Rental listing fraud detector. Paste a Craigslist URL or raw listing text — get back a trust score, red flags, and positive location signals.

---

Architecture

POST /analyze
      │
      ├─► scraper.py        fetch + parse listing from URL (Member B)
      │        │
      │        └─► title, price_usd, description, image_urls
      │
      ├─► llm_analysis.py   LLM scam detection (Member C)
      │        │
      │        └─► suspicion_score, red_flags, reasoning, accessibility_signals,
      │            formatted_description
      │
      └─► main.py           orchestration + response (Member A)
               │
               └─► trust_score, verdict, red_flags, llm_score, price_score,
                   accessibility_signals, market_price_score, neighborhood_note,
                   neighborhood_info, listing_title, listing_description,
                   listing_price_usd, listing_image_urls

Trust score formula:

trust_score = 100 - ((llm_score * 0.6) + (price_score * 0.4))

---

Team

Member	File	Responsibility
A	main.py	FastAPI server, orchestration, scoring
B	scraper.py	Craigslist scraper (httpx + BeautifulSoup)
C	llm_analysis.py	LLM fraud analysis (OpenAI primary, Anthropic fallback)
D	frontend/index.html	Vanilla JS/HTML UI

---

Features

• Trust score (0–100) — weighted composite of LLM analysis + price heuristic
• Red flags — short, plain-English fraud signals extracted by the LLM
• Accessibility signals — positive location signals extracted from listing text (transit, ADA, parking, walkability, elevator)
• Listing preview — scraped title, price, and paraphrased description shown in a clean card above the result; image gallery with thumbnails
• Reverse image search — Google Lens + TinEye links under each listing photo; no API key needed
• Paraphrased descriptions — LLM rewrites raw Craigslist text (ALL CAPS, separator lines, emoji spam) into clean professional prose
• Clickable history URLs — listing URLs in the history panel are hyperlinked back to the original listing page
• Demo mode — keyword-based frontend analysis with no backend required
• Analysis history — last 50 analyses stored in localStorage
• Dual LLM — OpenAI gpt-4o-mini primary, Anthropic claude-haiku-4-5 fallback

---

Setup

1. Clone and enter the repo

git clone https://github.com/NilayRaut/rentsentry.git
cd rentsentry

2. Create the conda environment

conda create -n rentsentry python=3.11 -y
conda activate rentsentry

3. Install dependencies

pip install -r requirements.txt anthropic httpx beautifulsoup4 lxml

4. Add your API keys

cp .env.example .env   # or create .env manually

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...   # optional — used as fallback only

---

Running

Start the API server

conda activate rentsentry
uvicorn main:app --reload

Open frontend/index.html directly in a browser — no build step needed.

Test with curl

curl -X POST http://localhost:8000/analyze \
  -H 'Content-Type: application/json' \
  -d '{
    "url": null,
    "description": "Wire transfer only, I am overseas on a missionary trip",
    "price_usd": 600,
    "image_urls": []
  }'

Expected response shape

{
  "trust_score": 10,
  "verdict": "likely_scam",
  "red_flags": ["Payment via wire transfer", "Landlord claims to be overseas"],
  "llm_score": 88,
  "price_score": 70,
  "accessibility_signals": [],
  "listing_title": null,
  "listing_description": "Wire transfer payment required. Landlord states they are currently overseas.",
  "listing_price_usd": 600,
  "listing_image_urls": []
}

---

API Contract

POST /analyze

Request

Field	Type	Default	Notes
url	string | null	null	Craigslist listing URL
title	string | null	null	Listing title
description	string | null	null	Listing body text
price_usd	number | null	null	Monthly rent
image_urls	string[]	[]	Always an array
mode	"rent" | "buy"	"rent"	Analysis mode

Response

Field	Type	Notes
trust_score	integer 0–100	Higher = safer
verdict	"safe" | "suspicious" | "likely_scam"
red_flags	string[]	Always an array, never null
llm_score	integer 0–100	Higher = more suspicious
price_score	integer 0–100	Higher = more suspicious
accessibility_signals	string[]	Positive location signals; empty if none found
mode	"rent" | "buy"	Echo of requested mode
market_price_score	integer 0–100 | null	Market price comparison; null if neighborhood unknown
neighborhood_note	string | null	Human-readable price context
neighborhood_info	object | null	Nearby amenities from LLM
listing_title	string | null	Scraped or provided title
listing_description	string | null	LLM-paraphrased description (falls back to raw)
listing_price_usd	number | null	Scraped or provided price
listing_image_urls	string[]	Scraped images; always an array
meta	object | null	Debug metadata (scrape attempted, price source, etc.)

---

Module: llm_analysis.py

from llm_analysis import analyze_listing

result = await analyze_listing(title, price_usd, description)
# {
#   "suspicion_score": int,            # 0-100
#   "red_flags": list[str],
#   "reasoning": str,
#   "accessibility_signals": list[str], # positive location signals
#   "formatted_description": str | None  # LLM-cleaned description prose
# }

• Primary model: gpt-4o-mini (OpenAI)
• Fallback model: claude-haiku-4-5 (Anthropic)
• Never raises — returns a safe default dict on any failure
• Run self-tests: python llm_analysis.py

---

Module: scraper.py

from scraper import scrape

data = await scrape("https://boston.craigslist.org/...")
# {
#   "title": str | None,
#   "description": str | None,
#   "price_usd": float | None,
#   "image_urls": list[str],
#   "phones_found": list[str]
# }

• Strips Craigslist QR footer from description automatically
• Test directly: python scraper.py (prompts for a URL)

---

Suspicion Score Guide

Score	Meaning
0–20	Looks legitimate
21–40	Minor concerns
41–65	Suspicious — verify carefully
66–85	Likely scam
86–100	Almost certainly a scam

---

Verdict Thresholds

Trust Score	Verdict
70–100	safe
40–69	suspicious
0–39	likely_scam

---

Future Scope

• More listing platforms — Facebook Marketplace, Apartments.com, Zillow (currently blocked by anti-scraping), HotPads; each needs a site-specific parser in scraper.py
• More cities / regions — price heuristics and neighborhood data are currently Boston-only; extend _BOSTON_MEDIANS pattern to other metro areas
• Image fraud detection — send listing photos to a vision model to flag stock images, digitally altered photos, or mismatched locations
• Phone number lookup — cross-reference scraped phone numbers against known scam databases (e.g. 800notes)
• Landlord reputation layer — crowdsourced reports tied to phone number, email, or address; flag repeat offenders
• Cloud history sync — replace localStorage with a user account so history persists across devices
• PDF / shareable report — one-click export of the full analysis as a PDF or shareable link
• Email / push alerts — notify a user when a saved search URL changes trust score or gets flagged

---

Browser Extension

A Firefox extension that injects an inline trust badge on Craigslist and Facebook Marketplace listing pages — no copy-paste required.

extension/
├── manifest.json          MV3 — permissions, content scripts, background
├── background.js          Service worker — proxies POST /analyze
├── content/
│   ├── craigslist.js      Extracts listing from static DOM (#titletextonly, .price, #postingbody)
│   └── facebook.js        MutationObserver for React-rendered pages
├── shared/
│   └── badge.js           Floating side panel UI (responsive: right-side on desktop, bottom bar on mobile)
└── icons/
    ├── icon16.png
    ├── icon48.png
    └── icon128.png

Load in Firefox (temporary, for testing)

1. Go to about:debugging → This Firefox → Load Temporary Add-on
2. Select extension/manifest.json
3. Open any Craigslist listing — the RentSentry panel appears on the right within a few seconds

Package for AMO submission

cd extension && zip -r ../rentsentry-extension.zip . && cd ..

Then upload rentsentry-extension.zip at addons.mozilla.org.

Requirements

• Firefox 109+ (Manifest V3 support)
• Backend must be running (update DEFAULT_API in extension/background.js to your Render URL)
• A privacy policy is required for AMO submission — listing text (title, description, price) is sent to the RentSentry backend for fraud analysis; no data is stored