James Yamazaki

Built the C++ inference stack for Sumsub's NIST FRTE 1:1 submission, ranked 29th of 444 evaluated entries on the Mugshot track. Deployed ONNX-based inference and vector search over 100M+ embeddings.

Shipped fill-in-the-middle code completion for VS Code and JetBrains IDEs. Implemented repository-aware context retrieval using AST- and path-based indexing and Tree-sitter symbol extraction.

Served image classifiers for a product supporting 10k+ concurrent users with HPA and Triton Inference Server. Converted PyTorch models to TensorRT with layer fusion and INT8 quantization, reducing cloud inference cost by 23%.

Co-authored The Poison of Alignment, showing that removing refusal-style responses from SFT data improved MMLU, BBH, HumanEval, and DROP performance by 4-33%. Built and operated multi-node training infrastructure for LLM experiments on 64 H100 GPUs.

Led research on unsupervised anomaly detection in 3D CT using diffusion models and GAN priors, improving Dice by 20% relative to baseline. Built and owned the research team's MLOps stack, reducing deployments from hours to minutes.

Built an object-detection system for musical symbols using MMDetection, improving mAP@50 by 0.20 absolute over baseline. Deployed a production-ready inference service on Amazon SageMaker, reducing cost by 60% versus previous EC2 deployment.

Pursued a Bachelor of Science in Computer Science, graduating in 2022.