At Accurascan, I designed and deployed a face and document presentation-attack detection system combining liveness classifiers with YOLO-based document detection and segmentation. It passed iBeta Level 2 evaluation, and its inference pipeline delivers under-200 ms response time across NVIDIA GPU servers and edge devices.
I built evaluation protocols using subject-, card-, and data-source-separated splits to assess cross-domain generalization. I also optimized inference with quantization, pruning, ONNX/TensorRT conversion, and FP16/FP8.
Previously, at Legaltech Indi, I developed and fine-tuned transformer-based NLP models for conversational AI, text classification, document understanding, and information extraction. In medical imaging projects, I developed pipelines for classifying diseases from chest X-rays and detecting brain tumors in MRI scans.
