Behind the build - June 15, 2026

Building HomeSignal: Day One and an Initial Architecture for Property Intelligence

Our first working session produced a clear direction: a multi-model AI pipeline that describes what it can actually see in a property image, scores its own confidence, and routes low-confidence records to human review.

Paul Huber Paul Huber Co-Founder, HomeSignal

Something I didn't fully anticipate about starting a company is how quickly a good team cuts through the hard problems. When experienced people come at the same question from different angles, the conversation moves fast.

This week the HomeSignal team held our first working session. We walked out with a clear direction, a practical architecture, and a shared conviction that AI can surface genuinely useful information about residential properties.

Our goal is straightforward: use modern computer vision and AI models to extract meaningful, structured information from property imagery. Rather than relying on a single model to make all the calls, we're designing a system that combines multiple AI systems, confidence scoring, and human verification to improve accuracy and trust.

1. Architecture

Why multiple models?

One of the core challenges with AI-powered image analysis is that every model has strengths and weaknesses. Some are excellent at detecting visible features. Others are better at reasoning about context. Rather than treating any single model as the source of truth, we're designing a system that lets multiple models evaluate the same property images independently.

The pipeline looks like this:

  1. Property images
  2. Vision analysis — pass one
  3. Vision analysis — pass two
  4. Cross-validation and scoring
  5. Frontier model evaluation
  6. Structured property features
  7. Confidence threshold check
  8. Human review (if needed)

Multiple vision models assess the same images and generate structured observations. Those observations are compared, scored, and reviewed by a larger reasoning model acting as a judge. When confidence is high, the system produces structured output automatically. When confidence falls below a defined threshold, the record is flagged for human verification.

2. Principles

Focusing on evidence instead of guessing

One of the principles we spent the most time on is that AI should describe what it can actually observe — not what it suspects or infers beyond that.

A property image may show evidence of deferred maintenance, roof wear, landscaping condition, exterior damage, or renovation quality. But there are plenty of things an image simply cannot tell you reliably.

Our system will explicitly distinguish between:

  • Visible evidence — something clearly present in the image
  • Reasonable inference — a conclusion supported by what's visible
  • Unknown or not observable — things the image can't answer

That distinction matters. The goal isn't to replace a professional inspection. The goal is to generate reliable property signals that can help inform decisions.

3. Technology

Why Gemini?

For our first prototype, we're planning to use Google's Gemini vision capabilities as the foundation of the image analysis pipeline. Gemini provides strong multimodal capabilities and clean structured output options, making it a good platform for rapid experimentation. Starting with one foundation model lets us validate the architecture before expanding to additional providers.

The objective of the prototype is not perfection.

The objective is to learn. Specifically, we want to understand:

  • Which property features can be extracted reliably
  • Where models agree and where they don't
  • How confidence scores correlate with actual accuracy
  • What types of images produce the best results
  • When human review is genuinely needed

4. Team

What I'm most excited about

Honestly, it's the team.

Collectively, we bring decades of experience across real estate, data platforms, cloud architecture, search technologies, machine learning, and software engineering. The conversations on day one moved past "Can AI do this?" almost immediately and into "How do we build a system we can trust?"

That's the question that separates interesting technology from useful products.

There is a lot to learn and a lot to build. But after one session, I'm more convinced than ever that there's a real opportunity here — practical tools that help people better understand the condition and characteristics of residential properties.

I'll keep sharing what we learn as HomeSignal develops.

Learn more about what HomeSignal is building, meet the people building it, or browse more HomeSignal field notes.