DSML contribution
Meetup speakers
Share their experience at our events
James Yamazaki
@yamazakiConnected with DSML. This person's Telegram account has been confirmed through DSML Kazakhstan.
Post training/Research Engineering/ML systems
ambiguous person, love to stare at graphs and loaded gpus, love reinforcement learning and reinforcement learning
Featured
Career
Experience: 7 yearsExperience
Machine Learning Engineer
Sumsub
Identity verification platform (KYC/AML)
Sep 2024 — PresentBuilt and deployed a rights-cleared face-recognition stack, including GDPR data controls and search across 100M+ embeddings.
• Built an EU-compliant face recognition stack end-to-end: license-audited dataset with provenance tracking, training, evaluation, and deployment. • Matched InsightFace antelopev2 performance while keeping the dataset fully rights-cleared. • Productionized via ONNX and a vector database with 100M+ embeddings for scalable search. • Implemented audit-ready data lineage and GDPR retention controls (consent, revocation, and per-sample purge), plus region-aware storage routing.
Research Engineer
Zencoder.ai
Startup building agentic coding tools and IDE integrations
Jan 2024 — Aug 2024Shipped IDE code-completion and repository-context features, and reduced end-to-first-token latency by 30%.
• Shipped a code-completion module (FIM/infill) for multi-language IDE extensions with token streaming, speculative decoding, and n-best ranking. • Implemented repository-aware context retrieval: AST- and path-based indexing (tree-sitter), a cross-repo symbol graph, and sliding-window prompt assembly. • Built an offline evaluation harness for completion acceptance (HumanEval/SWE-Bench-style tasks, exact match, and edit distance) and online telemetry for acceptance rate and dwell time.
ML/Inference Engineer
INUI Gaming
Gaming marketplace; large-scale image classification in production
Aug 2023 — Jan 2024Productionized image classifiers for 10k+ concurrent users and cut cloud costs by 23% through inference and CPU-path optimization.
• Productionized image classifiers for 10k+ concurrent users with HPA and Triton Inference Server. • Converted PyTorch models to TensorRT with layer fusion and per-channel quantization; introduced dynamic batching and request coalescing. • Optimized hot-path CPU bottlenecks in preprocessing and postprocessing, cutting cloud costs by 23%.
Research Engineer
GOAT.AI
LLM research
Jun 2023 — Aug 2023Co-authored a paper on alignment objectives in model distillation and managed GPU-cluster and distributed-training infrastructure.
• Co-authored The Poison of Alignment: an information-theoretic analysis showing alignment objectives contaminate student models during distillation; open-sourced artifacts. • Was mainly responsible for training infrastructure, including GPU cluster management and distributed training.
Computer Vision Engineer
Cerebra
MedTech startup helping in early detection of stroke
Oct 2022 — Jun 2023Developed medical-imaging models and data infrastructure, with reported improvements of 20% in Dice, 23% in IoU, and 1.8x processing speed.
• Developed unsupervised anomaly detection using DDPM and GAN priors; improved Dice by 20% on an internal target. • Built 3D CT hemorrhagic-stroke segmentation, improving IoU by 23% over baseline. • Built a medical-data processing framework 1.8x faster than standard preprocessing. • Introduced MLOps (CI/CD, model and data governance) to shorten time to production and improve tracking.
Skills
PythonC++JuliaPyTorchJAXLightningONNXTensorRTCUTLASSCUDADockerUnixAWSGCP