OPEN TO WORK · ML / LLM / Bioinformatics

Building private AI &
LLMs, in-house.

Software Engineer / Machine Learning · Local LLMs · Bioinformatics

Engineer with a MITOU (未踏) project selection and NeurIPS / ICML Workshop papers. I design and build local LLM platforms for fully air-gapped, high-security environments, lead AI adoption in contact centers, and work across bioinformatics — owning the whole path from data pipeline and model optimization to production operation.

View experience → Tech notes (Zenn)
6
Major projects delivered
7+
Peer-reviewed / workshop papers
100+
Concurrent users at 1.5s latency
500h
Monthly hours saved (DX unit)
// about

Profile

After graduating, I researched machine learning models and their application to healthcare data in graduate school, with papers accepted at NeurIPS and ICML workshops. Over roughly six years I have led development and operations across six projects.

Beyond large-scale transformers for NLP and vision, my current focus is fully air-gapped local LLM platforms and AI adoption in contact centers. I take responsibility all the way to real-world usage — not just building a model, but making sure it is adopted on the floor. Bioinformatics — cancer heterogeneity and RNA-seq analysis — remains a core strength backed by publications.

// expertise

Expertise

Research-grade theory paired with implementation under strict on-prem, air-gapped constraints.

🔒

Local LLM / Air-gapped AI

Design, build and optimize on-prem LLM stacks that never touch an external API — privacy-preserving RAG and inference acceleration.

Llama 3.1Mistral Large 2vLLMTensorRT-LLMSFT / DPOFP8 / AWQ
🧬

Bioinformatics

Cancer heterogeneity prediction & visualization, tumor-microenvironment analysis from RNA-seq, and ML on biomedical data.

RNA-seqSingle-cellCancer GenomicsSelf-SupervisedMedical Imaging
🎧

Contact-center AI

Inquiry automation, agent-assist, summarization and knowledge search — production-grade RAG with proper evaluation.

RAGVoice / DialogueFAQ automationAgent assist
🤖

Machine Learning / NLP

Large transformers for NLP and vision. Anomaly detection, embedding search, self-supervised learning.

PyTorchTransformersLoRAEmbedding Search
⚙️

MLOps / Infrastructure

In-house from GPU server selection to deployment. CUDA/Docker, quantization, inference tuning, vector DB ops.

A100 / H100 / L40SCUDADockerMilvus
🧭

AI-driven dev / PM

PM for teams up to 10. Established AI-first workflows, drove organizational adoption, and built quality-assurance processes.

AI-assisted devCursorPrompt designReview design
// experience

Experience

Most recent first. Figures are outcomes delivered in each project.

2025.09 – 2026.06 · 5 moLatest

Local LLM for personal & confidential data

Major manufacturer (1000+ employees) · 40 person-months / 5 in-house / 1 vendor

  • In-house infra: selected on-prem GPU servers with NVIDIA A100/H100/L40S; set up CUDA/Docker.
  • Model optimization: accelerated Llama 3.1 (70B/405B) and Mistral Large 2 with vLLM / TensorRT-LLM.
  • Domain tuning: anonymized confidential docs and PII logs, then fine-tuned via SFT and DPO.
  • Privacy-preserving RAG: built Milvus inside the air-gapped network with ACL-based authorization.
  • Quantization: FP8 / AWQ tripled inference throughput while preserving accuracy.
Zero external egress · 100% policy compliance +60% accuracy on in-house jargon 1.5s avg latency at 100 concurrent users 500h/mo saved in DX unit
Embedding AI-driven development in the organization

Shifting a "hand-coded + human-reviewed" culture to AI-first

Veteran engineers rejected and distrusted AI-generated code

Demonstrated legacy-code analysis/refactoring and bulk boilerplate generation with AI, showing the speed gap in numbers. Defined an AI collaboration guideline — "use AI as a junior, humans focus on architecture" — to drive a mindset shift.

AI made design work hollow and degraded code quality

Made a "precise prompt spec for the AI to implement against" a deliverable of detailed design, and added a pre-coding review step that uses an LLM to detect spec contradictions before Cursor generation — keeping speed while assuring quality.

2024.06 – 2025.01 · 8 mo

ML-based blockchain analytics business (Hong Kong)

Crypto-asset transaction monitoring · 85 person-months / 9 in-house / 3 vendors

  • Designed an ML anomaly-detection system flagging suspicious transactions in real time.
  • Built network-analysis tooling to visualize relationships between wallet addresses.
  • Implemented risk scoring from transaction history and automated smart-contract security risk detection.
+35% detection accuracyUncovered large fraud networks, worked with law enforcement
2023.01 – 2024.05 · 14 mo

Privacy-focused local LLM

Cajal Inc. · insurance & healthcare · 60 person-months / 6 in-house / 2 vendors

  • Evaluated Llama / Gemma / Qwen, selected Llama-3-8B-Instruct; LoRA fine-tuning with PyTorch + Hugging Face.
  • Curated medical-terminology datasets; optimized inference on constrained devices via Llama.cpp quantization.
  • Built medical-document summarization and QA; deployed to hospitals with no external data egress.
Cut specialist review timeSuccessful hospital deployment
2022.01 – 2023.12 · 12 mo

AI-powered personalized marketing platform

Cajal Inc. · 70 person-months / 7 in-house / 3 vendors

  • Built a real-time ad-targeting AI engine and customer-segmentation models.
  • NLP-based social listening, automated A/B testing, edge-AI delivery. GDPR/CCPA-compliant design.
CTR +25%CVR +30%Analysis effort -40%
2021.01 – 2021.12 · 12 mo

Legal-document search with large transformers

Cajal Inc. · 75 person-months / 7 in-house / 2 vendors

  • Built search algorithms on BERT / GPT-3 and a law-specialized BERT; developed a QA-style search system.
  • Created embedding models for similar-precedent retrieval; on-prem, security-conscious contract risk analysis.
+20% search accuracy-50% reading time+30% search efficiency
// publications

Papers / Bioinformatics

Machine learning × biomedicine — cancer heterogeneity, clinical text, and medical imaging.

2017T. Nakano, K. Ikeda, “Predicting Cancer Heterogeneity from One-shot Biopsy,” ICML Workshop on Computational Biology.PDF
2017T. Nakano, K. Ikeda, “Visualizing Cancer Heterogeneity with Dynamic Flow,” ICML Workshop on Computational Biology.PDF
2016R. Sato, T. Nakano, et al., “RNA Sequencing Analysis Reveals Interactions between Breast Cancer or Melanoma Cells and the Tissue Microenvironment during Brain Metastasis,” BioMed Research International.PDF
2016T. Nakano, “Generating Clinical Texts from Conversation,” NIPS Workshop on Machine Learning for Health.PDF
2015M. Hashimoto, T. Nakano, H. Yashiro, T. Fujita, M. Jinzaki, “Application of Word2vec to Radiology Report Analysis,” Japan Association for Medical Informatics, vol. 35.PDF
2015T. Nakano, “Early Detection of Hepatorenal Syndrome from Medical Records,” NIPS Clinical Workshop.PDF
2014T. Nakano, “Unsupervised Feature Detection for Medical Images,” 28th Annual Conference of JSAI.
// education

Education

Keio University, School of Medicine — B.M.
2017.03
The University of Tokyo, Graduate School of Medicine
2021.03 · withdrew