Inspiration
This project was inspired by the real-estate investment background on our team. Investors often “drive for dollars,” physically driving through neighborhoods looking for distressed or run-down properties that might signal a motivated seller. While effective, it’s slow, location-bound, and impossible to scale. We wanted to automate that entire process using computer vision and LLM-based analysis so that any market could be evaluated instantly, without ever getting in a car.
What it does
Revealio automates property distress detection using Google Street View imagery and multi-agent AI workflows. It systematically captures exterior images of every property in a target market, grades each home’s condition based on multiple weighted factors, and generates a distress score. The system then compiles the score with property-level data, ownership information, mailing addresses, and public records, into a refined, high-value marketing list.
Instead of buying broad, unfiltered city-wide lists, investors get a precise, data-driven list of the most promising off-market opportunities.
How we built it
-Integrated with Google Maps / Street View APIs to capture property-specific exterior imagery at scale. -Built a multi-task, multi-agent LLM framework using AKKA, RedPanda, OpenAI and Anthropic. -Configured two complementary grading LLMs, one evaluating distress indicators and the other evaluating the opposite, allowing us to detect disagreements, improve accuracy, and reduce false positives. -Designed the pipeline to compile condition scores with ownership and property-record data (via CoreLogic API). -Architected everything to run at scale so entire cities can be processed in minutes.
Challenges we ran into
-CoreLogic API failure: The property-data provider we secured access to had a platform glitch and couldn't generate API keys at the last minute, preventing live ownership data pulls. -Street View angle issues: Extracting consistent and relevant imagery for each parcel required more engineering than expected because Google sometimes returns odd or partial angles. -Agent consensus: Getting two LLM models to disagree constructively, and then to reconcile their reasoning, took significant iteration. -Time constraints: Multi-agent workflows + real geospatial data meant rapid debugging and tight coordination.
Accomplishments that we're proud of
-Successfully created an automated, scalable alternative to “driving for dollars.” -Built a functional multi-agent LLM grading system that cross-checks itself to improve accuracy. -Integrated computer-vision-driven property condition analysis with geospatial data collection.
What we learned
-How to orchestrate multi-agent systems where LLMs perform different roles and verify each other’s work. -How to interact with Google Maps APIs to systematically capture property-level imagery. -The complexities of matching geospatial data, imagery, and property-record data in one streamlined workflow. -That scaling real-estate lead generation is far more achievable with AI than traditional methods.
What's next for Revealio - Automated Computer Vision for Property Condition
-Full CoreLogic integration once their API issue is resolved, enabling live ownership and contact data. -Automated outbound campaigns (mail, text, email) directly triggered by distress scoring. -Batch-level API so wholesalers, flippers, and institutional buyers can pull full-city analyses on demand.
Built With
- akka
- anthropic
- corelogic
- docker
- google-maps
- loveable
- neo4j
- openai
- python
- redpanda
Log in or sign up for Devpost to join the conversation.