OpenHouse

The Problem

Every property listed online is still experienced as a flat grid of disconnected photographs. You can see what a kitchen looks like, but you can't feel how it connects to the dining room, how the light moves through the space at midday, or whether the balcony is actually worth a forty-minute drive to view.

The real estate industry has known about 3D tours for over a decade. Matterport exists. The technology isn't new. And yet the overwhelming majority of listings, globally, are still published as disconnected JPEGs.

Why? Because producing a 3D walkthrough requires specialised LiDAR equipment, trained operators, post-processing software, and hours of manual QA. For a typical agent managing 30 active listings, it simply isn't viable.

We built OpenHouse to remove that barrier entirely.

What OpenHouse Does

OpenHouse is an autonomous multimodal agent that intercepts a newly published property listing and handles everything required to produce a verified, interactive 3D experience, with no manual 3D work from the agent or realtor.

The pipeline runs like this:

Intercept — When new listing is published and then OpenHouse detects the event automatically via webhook and begins working immediately.
Analyse — Gemini 3.7 Flash evaluates the listing metadata, available media, and floor plan. It identifies advertised rooms and assesses spatial coverage against capture quality.
Recapture — If a spatial gap is found (a missing transition, an uncaptured room), OpenHouse drafts an exact 15-second recapture request. The on-site agent receives a secure, zero-install mobile link via WhatsApp. The mobile interface guides their physical movement frame by frame.
Verify — Submitted footage is instantly verified by Gemini for spatial continuity before the pipeline resumes.
Build — Camera poses are aligned, spatial evidence is aggregated, and the interactive experience is assembled.
Publish — The realtor receives a Confidence Ledger, not vague percentages, but verifiable spatial claims — and approves with one click. The 3D experience goes live on the web, instantly accessible on any browser, on any device, with no app required.

What Makes It Genuinely Agentic

Most "AI tools" in real estate act as smarter filters, they still require a human to input clean data, select settings, and interpret results.

OpenHouse operates differently. When Gemini flags that the outdoor pool-to-guest-house transition is missing from the walkthrough footage, the system doesn't just highlight the problem and stop. It:

Identifies exactly what spatial continuity was broken and why it matters for buyer navigation
Drafts a precise, spatially-aware recapture brief with directional instructions
Dispatches it to the appropriate on-site contact with a secure one-time mobile link
Verifies the returned footage for spatial correctness before resuming
Handles the full recovery autonomously

No human had to write a brief, open a 3D editor, or re-run a process. The agent took ownership of the entire remediation loop.

The Buyer Experience

The final output isn't a static panorama or a stitched photo sphere. It's an interactive, browser-native 3D experience that:

Loads instantly on any phone or laptop, no app, no plugin
Offers both a guided one-minute walkthrough and free spatial exploration
Includes a spatial AI assistant that answers buyer questions grounded strictly in verified property evidence, never hallucinating measurements or features it can't confirm
Surfaces a one-tap booking flow so only genuinely interested buyers request in-person viewings

The 3D experience doesn't replace the in-person viewing. It ensures only serious buyers show up.

How We Built It

The core architecture is an event-driven state machine that manages the full property lifecycle — from listing detection through spatial QA, recapture dispatch, pipeline resumption, and live publication.

Gemini 3.7 Flash (via the Google GenAI SDK) powers:

Multimodal spatial analysis of listing photos and floor plans
Spatial gap detection and recapture brief generation
Submitted footage verification for continuity before pipeline resumption
Grounded question-answering in the buyer-facing spatial assistant

Google Cloud Run hosts the agentic backend with autoscaling from zero, keeping costs proportional to volume. Cloud Build handles CI/CD.

The realtor dashboard, capture mobile interface, and buyer 3D viewer are built in React 18 + TypeScript + Tailwind CSS, deployed on Vercel.

Challenges We Faced

Agentic reliability over multi-step pipelines: The hardest engineering problem wasn't any single AI call, it was ensuring the state machine could recover gracefully from partial failures across a 6-stage pipeline. A spatial verification failure at step 4 needed to cleanly re-trigger step 3 without corrupting upstream state.

Grounded spatial AI: Getting Gemini to answer buyer questions accurately without hallucinating room dimensions or features required careful prompt design and strict evidence scoping. The assistant explicitly signals uncertainty when it can't verify a claim, which turned out to be one of the most trusted features in user testing.

Zero-friction mobile capture: Designing a mobile interface that could guide a non-technical on-site agent through precise spatial capture, with no app install, no login, and under 30 seconds from link-tap to submission, required significant iteration on the guidance UX.

What We Learned

Giving an AI agent responsibility , not just capability, is the design shift that matters. OpenHouse doesn't assist a human through a 3D production workflow. It owns the outcome. The realtor's only decision is: "Does this meet the standard?" And that single approval click is all it takes.

The confidence ledger model (explicit, verifiable evidence claims rather than black-box quality scores) significantly increased realtor trust and reduced time-to-approval in our testing. Transparency about what the system genuinely cannot confirm was more effective than hiding uncertainty.