Realtogo

Lofty AI, the agent that works while you're busy being an agent. Lives in WhatsApp. No app. No login. Just deals.

Comment

Inspiration

Real estate agents do their best work in the field, but their tools are built like office software. They have to bounce between texts, CRM tasks, listing notes, follow-ups, and market research while driving, hosting open houses, or walking a showing. We wanted to build an ambient assistant that meets agents where they already are, inside WhatsApp, and acts on context before they have to stop and open another app.

What it does

Our AI agent is a WhatsApp-native AI field assistant for real estate agents. It supports five core workflows: Open House Live Mode: detects when an agent has been at a property long enough and switches into lead-response mode. Voice to Listing: turns a WhatsApp voice note into a polished listing description and matched-buyer follow-up. Driving Briefing: generates a hands-free voice briefing with overdue tasks and hot leads. Pocket Card: uses shared location to generate nearby market context and listing intelligence. Passive Listener: captures showing conversations, extracts promised follow-ups, and drafts both an agent recap and buyer follow-up.

How we built it

We built a Node.js + Express backend with Twilio WhatsApp webhooks as the main interface layer. We used Vertex AI / Gemini for reasoning, drafting, and structured extraction, plus voice transcription for inbound audio. ElevenLabs handled voice synthesis, Cloudflare R2 hosted audio files, and Supabase stored demo CRM data like profiles, leads, tasks, saved location context, and seeded listing inventory. We also built a React demo panel so judges can trigger flows live while still seeing the results arrive on a real phone. We also linked it up to Zillow for database searches.

Challenges we ran into

The biggest challenge was reliability across real-world messaging flows. WhatsApp voice notes do not always arrive with perfectly consistent metadata, so getting Passive Listener to correctly recognize and process inbound audio took multiple iterations. We also ran into Twilio sandbox delivery limits, which affected testing. Another challenge was making AI output feel polished and consistent, especially when dealing with partial transcripts or incomplete model responses.

Accomplishments that we're proud of

We are proud that this feels like a real product, not just a hackathon mockup. The assistant works inside WhatsApp, supports real voice-note flows, delivers audio back to the phone, remembers profile and location context, and handles multiple field workflows end to end. We are also proud that we moved beyond hardcoded demo logic into a more durable system with saved context, fallback handling, and Supabase-backed data.

What we learned

We definitely learned a lot regarding connecting API's and building a product with no proper interface. We talked to one of the judges and had a conversation about how technology is moving away from interfaces. Taking that into regard, we built an application that isn't reliant on a proper interface to take into the account the judges insight. It was our first time ever building something like this.

What's next for ARN

We would ideally love to connect it with Lofty's database and get the full setup working. We were limited with the free options we had to use but if we had more time and money, we can definitely make it better. We would add proper security and make it fully hand-off application (maybe it being triggered by Siri/Google).

Built With

Share this project:

Updates