KOEN
Contact
Home / Solutions

InfraSafe AI — reading the site with generated data

Four solutions built on the same generation engine. Where the domain changes, the engine does not — only the parameters do.

Sewer networks

SewerDefect AI

Reads defects from in-pipe CCTV footage and drafts the inspection report.

Municipal sewer inspection means watching endoscope footage end to end — tens of minutes per pipe, thousands of pipes per contract. The real problem is not time but consistency: two inspectors bound the same crack differently, and when staff change the criteria leave with them. SewerDefect AI removes the subtitle and overlay artefacts, generates pixel-precise polygons with SAM2, sorts them into three confidence tiers, and classifies the defect — so people review the middle tier instead of the whole set.

23
defect classes covered
74.4%
field recall, 4,520 GT images
83.6%
precision, 4,520 GT images
53,480
auto-generated precise labels
Input

Sewer inspection video (CCTV / endoscope)

Process

Overlay removal → SAM2 polygons → 3-tier confidence sorting → defect classification

Output

Defect type, location and grade · reading evidence · draft inspection report

Performance is measured against real ground truth, not reported as mAP. We use partial-overlap recall — did the defect get found — because that is the question the client is actually buying an answer to, and it lets buyer and supplier accept the work in the same language.
Subsurface safety

GroundSafe AI

Detects voids, leaks and ground loosening in GPR survey data and ranks drilling priority.

A void under a road is invisible until it becomes a sinkhole. Ground-penetrating radar leaves a signal, but telling a void from a buried pipe or a soil transition still depends on an experienced eye — close to four hours per survey section, and the drilling decision is a separate judgement after that. The anomalies that matter are rare, so training data is scarce.

≥85%
detection mAP@0.5 — program target
240 → 20 min
reading time, 90% reduction target
20,000
synthetic sets to be built
TRL 4 → 7
2026.09 – 2027.03
Input

GPR survey data · utility drawings · site logs

Process

GPR-DataSynOps synthesis → GPR-AutoLabel → anomaly detection model

Output

Anomaly location and type · drilling priority · evidence report

These figures are the targets of the K-water public-private open-innovation program, not achieved results. Verified outcomes will be published with the certified test report when the program closes.
Worker safety

WorkerSafe AI

Detects hazardous worker situations in real time in low-light and confined sites.

Site safety management is still mostly retrospective — footage reviewed after an incident. Real-time monitoring degrades sharply at night, underground and in confined spaces. The deeper problem is that collision, entrapment and fall events are things that must never happen, so the footage simply does not exist. Synthesis is the only way to build that training set.

RGB-IR
visible and thermal fusion for low light
Edge
on-device inference, minimal server load
Input

Site CCTV · RGB-IR cameras · site logs

Process

EventVideoSynOps hazard synthesis → RGB-IR fusion → edge-optimised model

Output

Real-time hazard alerts · auditable safety log

This is the solution where our generation engine is used most directly: the rarer and more unacceptable the event, the less real footage exists.
Built structures

StructureCrack AI

Detects cracks and surface defects in buildings, tunnels and bridges and grades severity.

Periodic inspection of buildings, tunnels and bridges is mandated by law. The inspection happens — but measuring crack widths, photographing them and writing the report is still manual, and it reproduces every problem we met in sewer networks: inter-inspector variance, reporting load, and no objective basis for maintenance priority.

Input

Inspection video · drone imagery

Process

Crack and surface defect detection → width and length quantification → severity grading

Output

Defect list · severity grade · draft inspection report

This extends a defect-reading structure already verified in sewer networks to adjacent structures. It is at planning stage — if you are evaluating it, tell us your site conditions first.
Contact

Tell us what has to be detected on your site

Tell us what has to be detected on your site. We will first assess whether that data can be generated. If you have sample footage, we will prepare an auto-labeling demonstration with it.

Request a demo or PoC See the technology