Triet Thompson

Designed and deployed a large-scale image similarity search service using PyTorch with ResNet embeddings and vector indexing, exposed through validated Python APIs to enable content matching across millions of images processed daily. Built a reusable model training package with dependency management and versioned configuration files for repeatable experiments and consistent artifacts. Implemented dataset preprocessing and augmentation pipelines with orchestration and background jobs to balance training data. Converted trained models to ONNX and optimized CPU inference graphs to reduce latency and stabilize throughput under concurrent load. Integrated TensorFlow Lite models with mobile teams through strict interface contracts and input normalization checks. Established CI/CD quality gates with unit testing of preprocessing and integration testing for inference endpoints, and configured monitoring alerts for latency and errors. Led rollback strategies to recover from drift and coordinate