Built a lightweight, sub-0.5s/image pipeline to read 16-digit card numbers and cardholder names from card photos using OpenCV (template matching) and Tesseract (OCR). On a small, hand-labeled 23-image set, the baseline achieves 48% recall on PAN and 65% on name. With feasibility of low-latency data capture (fraud/risk, checkout autofill) established, next steps include dataset scaling and custom OCR training to reach production readiness.
A practical, side-by-side walkthrough of two ways to stylize images: a custom VGG-19 approach you can fine-tune for unique brand looks, and a pre-trained TensorFlow Hub model that ships fast and scales easily. Includes links, code snippets, and the trade-offs product teams care about (control vs. speed, quality vs. effort).
Built and benchmarked a smartphone-first road-damage detector using YOLOv5 (with ensembling + test-time augmentation) and Faster R-CNN on the Global Road Damage Detection dataset. Achieved a top-5 leaderboard result (F1 0.68) across 121 teams while meeting a 0.5s/image inference target—enabling practical, low-cost deployment from dashboard-mounted phones. Includes a mapping concept (GPS → segment scores) to guide maintenance prioritization for DOTs and municipalities.
Led development of an intelligent video monitoring system that automatically detects moving objects in security footage. Successfully evaluated 8 different detection algorithms across 53 test videos, achieving 96% accuracy in ideal conditions and 82% in challenging lighting scenarios. This solution enables automated security monitoring, parking occupancy tracking, and retail analytics without requiring constant human oversight.
