AI/ML Engineer Portfolio

Python · PyTorch · Hugging Face · RAG · Docker · FastAPI

Open to mid-level AI/ML roles. Each folder has its own requirements.txt.

Japanese document Q&A (RAG)

Ask questions against Japanese company PDFs. The system has to retrieve the right passage and not invent figures.

• Japanese-aware chunking with overlap → bge-m3 → Chroma → FastAPI + Streamlit + Docker
• After the overlap rebuild, Rakuten’s Non-GAAP operating profit (1,063億円) and the “トリプル20” AI line rank first
• Extra rerankers were tried and left out — they pushed that profit figure down
• Folder: Japanese_RAG_Production

Credit card fraud

XGBoost on a heavily imbalanced set. Fraud recall 0.92, PR-AUC 0.85. SHAP: V14 / V17. Docker + tests.

Live demo · Folder

Japanese sentiment

Fine-tuned cl-tohoku/bert-base-japanese-v2 (positive / neutral / negative). CPU demo + model on Hugging Face.

Streamlit · HF model · Folder

Customer churn

Telco churn baseline. Accuracy 0.82, recall 0.57. Live form only.

Live demo · Folder