Lofty Atlas

Real estate agents spend 74% of their day on admin. Atlas gives Lofty a voice. It briefs you every morning, coaches you on live calls, and executes tasks so you can focus on closing.

Comment

Why We Built This

According to NAR, real estate agents spend only 26% of their time on actual revenue work. The other 74% is paperwork, follow-up, data entry, and coordination. And 87% don't survive their first 5 years.

It's not because they lack tools. Lofty has everything. Lead scoring, automated follow-ups, transaction tracking, six AI agents. The engine is incredible.

But we ran a live audit of Lofty's platform using Claude and Playwright. 48 screenshots. Every section. Here's what we found:

One agent who had been using Lofty for five years wrote on Capterra this January: "SmartPlans are very technical and cumbersome to set up. The AI assistant is robotic, unable to be instructed and unreliable."

And we saw it ourselves. 8 to 12 clicks before any real work on login. The AI sitting silently in a corner icon. Notifications with real activity while the AI said "nothing to see here." Six powerful AI agents buried two clicks deep that most agents never find.

Lofty even posted a YouTube Short in June 2025 called "Buried in task chaos?" — they were patching task filters. But the deeper problem, the AI having no presence, no voice, no reasoning — that patch didn't touch it.

The AI had everything except a voice.

What Atlas Does

Atlas is that voice.

Every morning when Baylee opens Lofty, Atlas speaks first. She already read the entire pipeline overnight. Three things that matter today, three reasons why, three actions ready to execute. No searching. No clicking. No guessing.

When Baylee calls a lead, Atlas coaches her in real time. Live talking points refresh on every conversation turn based on what the customer is saying. The agent focuses on the relationship. Atlas handles the context.

When the call ends, Atlas writes the summary, logs the full transcript, and saves everything to the database automatically.

And when Baylee wants anything else, she just asks. Add a lead from a pasted email. Build a Smart Plan in plain English. Draft a follow-up. Atlas knows every page of the platform and every piece of data in the pipeline.

How We Built It

Frontend: Next.js 15, React 18, TypeScript, Tailwind, Framer Motion

AI: Groq powers every generation path in the product. We use llama-3.3-70b for deep reasoning and llama-3.1-8b for live call coaching because speed matters more than anything when someone is on a phone call.

Voice: ElevenLabs handles two things. The Conversational AI WebSocket runs the live call simulation between Baylee and the lead. The TTS API reads out the reasoning behind every lead score so agents can actually trust what Atlas is telling them.

Backend: InsForge is our entire backend. Seven Postgres tables: leads, transactions, tasks, listings, appointments, smart plans, and call logs. Every call end writes a row, uploads the full transcript JSON to InsForge S3 storage, and links the CDN URL back to the record. We also use InsForge Edge Functions for the morning briefing cron — Atlas pre-computes Baylee's day at 6am before she opens the app. Real production backend, not a mock.

Input: Web Speech API handles mic input so Baylee can just talk.

What We Learned

Trust is everything. Agents have been burned by AI that gets things wrong. So Atlas never just gives a number. She explains it out loud, shows which CRM records she read, and lets the agent click through to verify. Every recommendation is grounded and auditable.

The other thing: the problem was never missing features. It was that a powerful platform had no front door a normal human could walk through.

Challenges

ElevenLabs WebRTC kept returning 404s so we forced WebSocket connections and built fallback logic for slow agent responses.

Groq at speed can hallucinate. We had to explicitly ban the model from inventing property details it could not verify from the database.

Getting InsForge Edge Functions wired for the morning briefing cron was the last thing we shipped — right under the wire.

AI Tools Used

Claude handled research, strategy, spec writing, and UX audit analysis.

Claude Code ran the live Playwright audit of Lofty's platform. 48 screenshots, full friction map, cross-referenced against real user reviews.

Groq runs all inference inside the product.

ElevenLabs powers the full voice layer, both conversational AI and TTS.

InsForge handles the entire backend — database, S3 storage, and Edge Functions.

Built With

  • claude
  • elevenlabs
  • groq
  • insforge
  • nextjs
  • react
  • tailwind
Share this project:

Updates