Pertumbuhan pesat sektor industri konstruksi baja menuntut penerapan standar kualitas yang tinggi, di mana kualitas sambungan las memegang peranan krusial. Inspeksi kualitas pengelasan baja di CV. Mulia Mandiri Konstruksi yang selama ini dilakukan secara manual rentan terhadap kelelahan manusia, inkonsistensi, dan memakan biaya operasional tinggi, sehingga diperlukan solusi otomatisasi. Penelitian ini bertujuan merancang dan membangun aplikasi deteksi cacat pengelasan baja berbasis Android menggunakan algoritma You Only Look Once (YOLO) guna mempermudah proses inspeksi. Metode pengembangan sistem memanfaatkan pendekatan Prototype, sedangkan proses identifikasi cacat menggunakan teknik instance segmentation model YOLOv8n-seg yang dilatih pada 1.800 citra sekunder (Good Welding, Undercut, dan Excess Weld). Hasil penelitian menunjukkan aplikasi berhasil diimplementasikan dan model mencapai performa Mean Average Precision (mAP50) Bounding Box sebesar 93,5% dan akurasi batas piksel (Mask) sebesar 92,8%. Evaluasi F1-score juga membuktikan ketepatan klasifikasi yang tinggi 100% untuk Good Welding, 97,17% untuk Excess, dan 73,85% untuk Undercut. Dengan demikian, sistem ini terbukti andal memetakan bentuk serta batas area cacat secara presisi, memberikan implikasi praktis berupa alat bantu inspeksi yang jauh lebih objektif, konsisten, dan efisien untuk industri konstruksi.
Kata Kunci: Baja, Cacat Pengelasan, Deteksi Objek, Prototype, You Only Look Once (YOLO).

The rapid growth of the steel construction industry demands high-quality standards, where weld joint quality plays a crucial role. Steel welding quality inspection at CV. Mulia Mandiri Konstruksi, conventionally conducted manually, is susceptible to human fatigue, inconsistencies, and high operational costs, thus requiring an automated solution. This study aims to design and develop an Android based steel weld defect detection application using the You Only Look Once (YOLO) algorithm to simplify the inspection process. The system development utilizes the Prototype method, while defect identification employs the instance segmentation technique of the YOLOv8n-seg model trained on 1,800 secondary images (Good Welding, Undercut, and Excess Weld). Results indicate the application was successfully implemented, achieving a Bounding Box Mean Average Precision (mAP50) of 93.5% and a pixel boundary accuracy (Mask) of 92.8%. Furthermore, the F1-score evaluation proves high classification accuracy 100% for Good Welding, 97.17% for Excess, and 73.85% for Undercut. Therefore, the system has proven reliable in precisely mapping defect shapes and boundaries, providing a practical implication as a significantly more objective, consistent, and efficient inspection tool for the construction industry.
Keywords: Steel, Welding Defects, Object Detection, Prototype, You Only Look Once (YOLO).