- Problem
- Listing copy and discovery had to scale across MLS-sourced inventory. Renters needed natural-language search; ops needed consistent descriptions for the public site, Facebook Marketplace packs, and other channels—without inventing fees, policies, or contact CTAs.
- Approach
- Shipped phrase rental search (LLM-ranked matches over published inventory, IP rate limits, graceful degradation). Built MLS description enhancement with channel-specific prompts (Gemini/OpenAI). Wired AI into marketplace posting packs and an agent-facing Miami prompt-library landing page. Grounded prompts with explicit “do not invent” rules for vouchers, HOA, and landlord policies.
- Outcome
- Live on vive.us: phrase search on home and /rentals, description enhancement in the MLS Supply Engine, marketplace packages, and SEO city pages—alongside Auth0, Prisma/Postgres, and Netlify production deploys.
- Lesson
- Marketplace AI needs a different contract per surface—ranking vs. copy vs. ops tools—and hard boundaries on what the model is allowed to assert.