Superb AI Builds Photo-Based Defect Sorting System for LH Housing

Superb AI, a Korean vision AI company, has built an automated defect classification system for Korea Land and Housing Corporation (LH). Tenants of LH rental apartments upload a photo of a cracked wall or a leaking ceiling, and the AI determines within one second which space, which material and what kind of defect is involved. The company has finished the first build phase and is now working on a second, upgraded phase.

A huge volume of complaints

LH receives roughly 300,000 defect complaint photos a month, or about 4 million a year. Until now, counselors sorted them one by one through text messages and phone calls. The classification scheme is complex: combining five criteria (space, part or material, defect type, trade, and construction type) yields more than 18,000 possible categories. Because counselors differ in experience, the same defect could be classified differently or wrongly.

One model instead of several

Superb AI handles the task with a single vision-language model (VLM), an AI model that understands images and language together. Systems typically use one model to locate a defect and another to identify its type. Here, one model classifies all five criteria from a single photo, which the company says simplifies the structure and reduces maintenance burden. Complaints are processed around the clock, and results come with a written explanation of why the model reached its judgment.

The system covers common defects such as cracks, leaks, damage, mold and stains, and infers likely causes from learned patterns even when a photo alone makes them hard to pin down. It was designed for detection accuracy of at least 90% and a false-positive and missed-detection rate of 10% or less. The company says it maintains that standard through regular monitoring and verification.

What changes for LH

Counselors can now focus on complaints that need human judgment rather than routine intake and sorting. Because defect data is now stored in a consistent format, LH can analyze which defects occur most often by housing complex, material and years since completion. That supports inspecting recurring problem sites in advance and setting replacement and repair timing for aging facilities based on data, as part of long-term maintenance planning.

CEO Kim Hyun-soo said the project solved the twin challenges of a complex classification scheme and large volumes of unstructured data with a single VLM. He added that the approach can extend to facility inspection and image-based complaint analysis across public services.