Clients have already paid for this
Not PoC credits — three consecutive years of delivery confirmed by contracts and tax invoices, across defense, environment, infrastructure and manufacturing.
Sewer defect auto-labeling
The AI draws the labels; people only check the ones worth checking
The client had assigned manual labeling of a sewer defect dataset to several engineers, and it failed. Given the same guideline, each person bounded cracks differently; drawing pixel-level polygons took minutes per image and fatigue compounded the error. Re-inspection reproduced the same variance at double the cost. With uneven label quality, model performance plateaued and the cause could not even be diagnosed.
Instead of adding annotators we changed three things. First, we unified the criteria, basing the defect taxonomy on AI Hub's public sewer dataset of over 410,000 images. Second, we generated the precise labels automatically, converting boxes into pixel-accurate outlines with the SAM2 foundation model — seconds instead of minutes. Third, we minimised review by sorting every generated label into high-confidence, needs-review and missed, so people inspect one tier rather than the whole set. The cost structure moves from full annotation to partial review.
Over 52,000 cleaned images and 53,480 auto-generated precise labels (box-fallback 0.27%). An auto-labeling pilot on 7,355 client images completed one full pass including confidence sorting. Verified against 4,520 ground-truth images, field recall was 74.4% and precision 83.6%.
The turning point was deciding not to report performance as mAP. mAP penalises an outline that is a few pixels off, but what a site needs is discovery — once a defect is found, grading and repair decisions are downstream steps. We proposed partial-overlap recall as the acceptance criterion, which put buyer and supplier on the same language and made contract negotiation considerably faster.
Defense drone AI video analytics
Aerial-view scarcity solved with a generation pipeline instead of flight hours
Footage looking down from a drone is nothing like ground-level footage: objects are small, angles extreme, backgrounds constantly changing. Public datasets do not carry the performance, and in the defense domain the real captures cannot leave the premises at all.
We applied AerialDataSynOps, our aerial-view object-detection data generation pipeline. Without any real capture, altitude, angle, illumination and season were combined as parameters to produce the scenarios the model was missing. Because no data leaves the site, no security issue arises.
A development contract worth KRW 100M (excl. VAT) delivered across two milestones. The same client placed a second order, making it a repeat customer.
Wildfire and algal bloom detection data
The pipeline built in year one was re-purchased in year two
Wildfires and algal blooms are monitored continuously but do not occur often enough to train on. The very intervals that need detection — the faint smoke of an early fire, a specific bloom concentration — are exactly where data is missing.
Two pipelines, FireDataSynForge and AlgaeDataSynForge, synthesised the target conditions, while existing captured footage was labeled.
Labeling and data generation were delivered in December 2025, and in August 2026 the client returned with an AI training-data enhancement contract. One pipeline, built once, served the follow-on demand in the same domain unchanged.
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.