SUSSY AI-POWERED RENTAL SCAM DETECTOR
6 specialist Claude agents run in parallel to score rental listings 0–100. Solo 5-day build — research to deploy. Live at sussy-ca.vercel.app, bilingual (EN / FR) from MVP.
Vancouver has one of the most competitive — and most scam-heavy — rental markets in Canada. A friend at UBC told me three people in her circle had lost money to rental scams. It clicked when I remembered my own first month in the city, trying to find a place and seeing the same red flags everywhere.
Build a tool students can use the moment they spot a sketchy listing — paste a URL or message, get a clear risk score, plain-English red and green flags, and what to do next.
A single AI prompt asked to check price, text, photos, and identity at once gets unfocused and misses subtle signals. JSON output breaks easily. Six parallel calls add latency and cost. The architecture had to be specialized, fast, fault-tolerant, and cheap to run — all without sacrificing accuracy.
Split detection into 6 specialist agents — each with one narrow job. All four analysis agents (Text, Price, Identity, Image) run in parallel via Promise.all, so total latency = the slowest agent, not the sum. An Orchestrator routes inputs and an Aggregator combines outputs into a weighted Risk Score 0–100.
- 01
Specialists beat generalists
Each agent has one job → sharper detection. A single prompt asked to check price, text, photos, and identity at once gets unfocused and misses subtle signals.
- 02
Parallel ≠ slower
All four analysis agents run via Promise.all. Total latency = the slowest agent, not the sum.
- 03
Fault-tolerant
If one agent fails (e.g. Image Agent on a text-only listing), the rest still run and the score still comes out. No single point of failure.
- 04
Cost-bounded
Per-agent max_tokens limits + a sliding window (last 10 turns) on the Chat Agent keep API costs predictable at scale.
Mapped Vancouver scam patterns — fake listings, deposits before viewing, stolen photos. Talked to students. Sketched the 6-agent architecture on paper before writing a line of code.
Designed the 6-agent system — Text, Price, Identity, Image analysis agents in parallel; Orchestrator routes inputs; Aggregator combines into a weighted score. Moved Claude API calls server-side via Vercel serverless functions on day 2 so the key never touches client code.
React + Vite frontend, Supabase backend with row-level security for community reports. Strict JSON schema in system prompts + response validation to keep AI output reliable. Sliding window (last 10 turns) on the Chat Agent to cap cost. Bilingual EN / FR from MVP for students across BC and Quebec.
Live on Vercel with env-var management for API keys. Roadmap Phase 3 adds Scammer Fingerprinting — Phone Tracking shipped (4+ shared phone numbers triggers a Repeat Scammer badge); Photo Trace and Address History coming next.