Dr. Kenji Nakamura
Research Engineer | Applied AI & Systems Research
📧 [email protected] | 📱 (555) 654-3210 | 🔗 linkedin.com/in/kenjinakamura | 💻 github.com/kenjinakamura | 📚 scholar.google.com/kenjinakamura
San Francisco, CA
Professional Summary
Research Engineer with a Ph.D. in Computer Science and 6+ years bridging fundamental research and production engineering. Expertise in large-scale ML systems, applied deep learning, and high-performance computing. Published 14 peer-reviewed papers with 900+ citations. Designed and deployed ML infrastructure serving 100M+ users at Meta AI Research. Passionate about turning research insights into real-world systems that matter.
Technical Skills
Machine Learning: Deep learning, reinforcement learning, large language models (LLMs), computer vision, self-supervised learning, model distillation
Frameworks: PyTorch, JAX, TensorFlow, Triton, ONNX, TensorRT, vLLM, HuggingFace
Systems & Infrastructure: Distributed training (FSDP, DeepSpeed, Megatron-LM), CUDA/GPU optimization, model serving, high-performance C++, Rust
Data & Compute: Apache Spark, Ray, HDFS, Slurm, Kubernetes, custom data pipelines at petabyte scale
Languages: Python, C++, CUDA, Rust, Bash
Research Tools: LaTeX, Weights & Biases, Hydra, DVC, Jupyter, Matplotlib, seaborn
Professional Experience
Research Engineer | Meta AI Research (FAIR) | Menlo Park, CA
August 2020 - Present
- Co-designed training infrastructure for OPT-175B language model, enabling distributed training across 992 A100 GPUs with 70% hardware utilization — published as technical report
- Built efficient inference engine for large-scale vision-language models, reducing serving cost by 3x and latency by 60% vs. baseline
- Developed novel data augmentation pipeline for self-supervised vision model that improved ImageNet top-1 accuracy by 1.8pp
- Led implementation of FSDP training recipe adopted by 12+ internal research teams as standard practice
- Collaborated on 6 published papers spanning LLMs, vision, and multimodal representation learning
- Mentored 4 PhD interns from Stanford, MIT, CMU, and Berkeley; 2 converted to full-time research engineers
- Open-sourced 3 tools (data pipeline, profiling library, evaluation harness) with 2,000+ combined GitHub stars
Research Scientist Intern | Google Brain | Mountain View, CA
May 2019 - August 2019
- Implemented novel attention mechanism for sparse transformer architecture, achieving 2x speedup on long-sequence tasks
- Ran large-scale ablation studies across 50+ model configurations using TPU pods
- Contributed implementation included in TensorFlow Research Models repository
Research Engineer | MIT Computer Science and Artificial Intelligence Lab (CSAIL) | Cambridge, MA
September 2015 - July 2020
- Designed and built robotic manipulation system integrating perception, planning, and control achieving 94% grasp success rate on unseen objects
- Developed reinforcement learning environment used by 15+ research groups worldwide (3,500+ GitHub stars)
- Implemented simulation-to-real transfer pipeline reducing real-world fine-tuning data requirements by 90%
- Managed compute cluster of 40 GPU nodes for 20-person research group
Education
Ph.D. in Computer Science (Robotics & Machine Learning)
Massachusetts Institute of Technology (MIT) | Cambridge, MA Graduated: August 2020 Dissertation: "Sample-Efficient Sim-to-Real Transfer for Robotic Manipulation via Structured Representations" Advisor: Prof. Leslie Kaelbling
Master of Engineering in EECS
Massachusetts Institute of Technology (MIT) | Cambridge, MA Graduated: June 2015 GPA: 4.8/5.0
Bachelor of Science in Computer Science
University of Tokyo | Tokyo, Japan Graduated: March 2013 Summa Cum Laude, GPA: 3.96/4.0
Selected Publications
-
Nakamura, K., et al. (2023). "Efficient Training of Large Language Models via Selective Gradient Checkpointing." NeurIPS. [Citations: 210]
-
Nakamura, K., Zhang, Y., & LeCun, Y. (2022). "Self-Supervised Pretraining for Vision-Language Alignment at Scale." CVPR (Oral). [Citations: 185]
-
Li, W., Nakamura, K., et al. (2022). "Sparse Mixture-of-Experts for Efficient Vision Transformers." ICLR. [Citations: 143]
-
Nakamura, K. & Kaelbling, L. (2020). "Sim-to-Real via Structured Latent Space Representations." ICRA (Best Paper Finalist). [Citations: 134]
-
Nakamura, K., et al. (2019). "Sample-Efficient Robot Learning through Meta-Adaptation." RSS. [Citations: 98]
Total Publications: 14 | Total Citations: 900+ | h-index: 9
Patents
- US11876543 – "Adaptive Gradient Checkpointing for Memory-Efficient Large Model Training" (Meta AI, 2023)
- US11543210 – "System and Method for Sim-to-Real Policy Transfer in Robotic Manipulation" (MIT, 2021)
Open Source Contributions
- torchprofile – PyTorch model profiling library (1,200 stars, 80+ contributors)
- simrobot-env – Robotic simulation RL environment (2,300 stars, used by 15+ universities)
- llm-eval-harness – Unified evaluation framework for LLMs (700 stars)
Invited Talks & Service
Invited Talks:
- "Scaling ML Systems: From Research to Production" – ICML Industry Track (2023)
- "Efficient LLM Inference" – Stanford ML Symposium (2022)
- "Sim-to-Real Transfer for Robotics" – CMU Robotics Seminar (2021)
Reviewing:
- Program Committee: NeurIPS (2021-2024), ICML (2021-2024), ICLR (2022-2024), CVPR (2022-2024)
- Outstanding Reviewer Award – NeurIPS (2022)
Awards & Honors
- NeurIPS Outstanding Paper Award (2023)
- ICRA Best Paper Finalist (2020)
- NSF Graduate Research Fellowship (2013-2016)
- MIT Presidential Fellowship (2013)
Additional Information
Languages: English (Fluent), Japanese (Native) Interests: Robotics, AI safety, competitive programming, Go (board game)
This resume was created using markdown and converted to PDF at MarkdownResume.app