Andriy Yeroshevych

Platform for automated visual quality control of aerospace tube stamping, labelling, and marking. A five-stage on-device CoreML cascade — detection, segmentation, classification, OCR, and shape recognition — guides inspectors stamp-by-stamp, validating each result against the paper router and registration data. All inference runs on-device with no network dependency; end-to-end per-stamp detection completes in under 5 seconds.

[Task: Automate and enhance the process of tube QC checking](https://www.twine.net/signin

![](https://www.twine.net/signin

AR mobile platform for spatial defect capture, measurement, and registration in industrial inspection.
Replace manual, paper-based QC inspection workflows in wind blade manufacturing with a spatial AR tool that captures defect location, dimensions, and documentation in one place.

[Task: Replace manual, paper-based QC inspection workflows in wind blade manufacturing with a spatial AR tool that captures defect location, dimensions, and documentation in one place.]
(https://www.twine.net/signin

![](https://www.twine.net/signin

8,000+ towers, 150,000+ antennas across the network. Operators constantly replace, remove, and add antennas. The single source of truth for any tower’s state is the tower itself. Towers are surveyed several times a year, photographed from multiple angles and positions. These photos are the raw data that must be processed to derive the real state of the network assets.

[Project: Create automatic system of accounting of antennas over all towers](https://www.twine.net/signin

The client managed all asset records in spreadsheets. While the primary goal was to determine which antennas are mounted on each tower and build a system for keeping that data current, getting there first required migrating the entire process into a unified platform. We built the ML Vision pipeline and the management system in parallel. Let’s start with the interesting part.

![](https://www.twine.net/signin