WalkThrough

Turn your phone into an AI home inspector that sees issues, estimates costs, and builds professional reports in real-time as you walk through a property.

Comment

Inspiration

Home inspections are a critical step in every real estate transaction, yet they remain stuck in the past — costing $300-$500 per inspection, requiring days to schedule, and producing reports
that arrive long after the walkthrough is over. We asked: what if AI could do this in real-time?

The Gemini Live API's ability to process simultaneous voice and video streams made this possible for the first time. A homebuyer, agent, or investor could simply walk through a property with
their phone, have a natural conversation with an AI inspector, and walk out with a professional assessment report — all in the time it takes to tour the home.

What it does

WalkThrough is a real-time AI home inspection agent. You open the app on your phone, point your camera, and start walking through a property. The AI:

  • Sees each room through your live camera feed and proactively identifies issues — water damage, cracks, mold, outdated wiring, foundation concerns, safety hazards
  • Talks with you naturally, guiding you room-by-room, asking you to pan the camera toward areas of concern
  • Rates every finding by severity (Critical / Major / Minor / Cosmetic)
  • Estimates repair costs using real market data via Google Search grounding
  • Generates a professional assessment report with captured images, room-by-room breakdowns, and a total cost summary

It's like having a seasoned home inspector in your pocket.

How we built it

WalkThrough is built on the Google Agent Development Kit (ADK) using the Gemini 2.5 Flash Live API for real-time multimodal processing.

Architecture:

  • Frontend: Next.js/React mobile-first PWA that streams camera video and microphone audio via WebSocket
  • Backend: FastAPI server running the ADK agent runtime with custom tools for saving findings, capturing snapshots, tracking progress, and generating reports
  • Data: Cloud Firestore stores room assessments and findings; Cloud Storage holds captured images
  • Deployment: Dockerized and deployed to Cloud Run with Terraform-managed infrastructure

Agent Tools: The AI agent has six purpose-built tools — save_room_finding, capture_snapshot, complete_room, get_progress, search_repair_costs (Google Search grounding), and generate_report —
giving it the ability to not just observe and talk, but take structured actions that produce a real deliverable.

Challenges we ran into

  • Balancing latency and quality: Streaming live video to an AI model while maintaining natural voice conversation required careful tuning of frame sampling rates and audio buffering
    • Cost estimation credibility: Making repair cost estimates feel trustworthy meant grounding them with Google Search data and always presenting ranges rather than single numbers
    • Mobile-first constraints: The entire experience runs in a mobile browser — managing camera permissions, WebSocket stability on cellular networks, and touch-friendly UI on small screens
      added real complexity
    • Solo timeline: Building a full-stack multimodal AI application in 15 days as a solo developer meant ruthless prioritization — MVP first, polish second

Accomplishments that we're proud of

  • Built a fully functional real-time multimodal AI agent that processes live video and voice simultaneously — not a pre-recorded demo, but a genuinely interactive experience
    • The agent proactively identifies issues without being asked — it sees water damage, cracks, and safety hazards on its own, mimicking how a real inspector works
    • End-to-end value delivery: walk in with a phone, walk out with a professional PDF report complete with captured images, severity ratings, and cost estimates - Achieved natural, hands-free interaction — the user never touches the screen during an inspection, making it practical for real-world use
    • Grounded cost estimates with live Google Search data so the numbers reflect actual market rates, not hallucinated figures
    • Deployed the entire stack on Google Cloud with Terraform-managed infrastructure, making it reproducible and production-ready
    • Solo-built in under two weeks while maintaining code quality and a polished user experience

What we learned

  • The Gemini Live API is remarkably capable at simultaneous voice + vision understanding — it can identify subtle issues like hairline cracks or water staining while maintaining a natural
    conversation
    • Google ADK's tool system makes it straightforward to give an AI agent real-world actions beyond just chatting
    • The biggest unlock for multimodal AI isn't the technology — it's finding use cases where voice + vision together are genuinely better than either alone. Home inspection is a perfect fit
      because your hands are busy, your eyes need guidance, and the context is inherently visual

What's next for WalkThrough

  • Multi-language support for international real estate markets
    • Historical comparison — re-inspect a property months later and diff the findings
    • Integration with MLS listings to auto-populate property details
    • Specialized modes for different inspection types (pre-purchase, insurance, rental move-in/move-out)

Built With

Share this project:

Updates

Submission history