P2P Lending App: Fraud Prevention

Computer Vision & Security

Engineered a document liveness detection system to mitigate fraud within a P2P lending platform. This system was designed to prevent users from using photos of screens or printed copies of documents during the KYC process.

• Model Development: Built a multiclass image classification model designed to distinguish between authentic KTP documents and fraudulent reproductions. I experimented with different architectures to find the best balance between accuracy and inference speed.
• Dataset Engineering: Prepared, curated, and augmented datasets to improve model robustness across various lighting and orientation conditions. I implemented custom augmentation pipelines to simulate common camera artifacts.
• Evaluation & Training: Utilized Google Colab for model training and iterative evaluation of accuracy and precision metrics. The model achieved over 95% accuracy on our validation set.

Taxonomy Design

• Class Definition: Defined a rigorous class taxonomy for liveness detection to handle nuanced fraud patterns, such as "screen-re-photography" vs "paper-re-photography".
• Integration Strategy: Worked closely with the mobile engineering team to design an efficient workflow for capturing high-quality document images while providing real-time feedback to the user.
• Security Impact By automating the liveness check, we reduced the manual verification workload by 40% and significantly lowered the rate of fraudulent account creations. This contributed to a safer lending ecosystem and increased investor confidence in the platform.