Software engineer — backend, search infrastructure, and full-stack products. Exploring applied AI and RAG.
I'm a software engineer who works mostly on the backend — search platforms, data pipelines, and full-stack products used by real users. At Akatsuki and Talendy I've built scraping-and-indexing pipelines behind a job-search platform and worked on an Azure → AWS search migration.
Lately I've been going deeper into applied AI — building RAG systems and, where it helped a project, some model fine-tuning. I like understanding what's under the framework, so I've built a RAG pipeline from scratch (no LangChain, no vector DB) to learn how retrieval and grounding actually work. My competitive-programming background (ICPC Asia West AIR 102, 1000+ problems) keeps me careful about correctness and edge cases.
Building with RAG · LangChain · Gemini
|
A team-built full-stack platform connecting restaurants' surplus food with NGOs. Real-time orders and chat over Socket.IO, hardened JWT-in-httpOnly-cookie auth, and a recommender built from scratch in NumPy. Stack: React, Flask, MongoDB, Socket.IO, NumPy |
A FastAPI service for finding jobs by meaning, not keywords — Gemini embeddings, a FAISS vector store, a LangGraph workflow, and cited answers grounded only in retrieved listings. Deployed with Docker. Stack: Python, FastAPI, LangChain, LangGraph, FAISS, Gemini, Docker |
|
A learning project: a retrieval-augmented pipeline over PDFs with no LangChain and no vector database, to understand how chunking, embeddings, retrieval and grounding actually work. Answers cite their source page and refuse questions the document doesn't cover, with a small eval harness (hit@k, MRR). Stack: Python, sentence-transformers, FastAPI, Docker, Gemini |
A university group project where I explored applied ML: fine-tuning DistilBERT for SMS smishing and comparing classical models (Logistic Regression vs Random Forest) for URL classification. Ships with a Chrome MV3 extension that scans page links against the deployed model. Stack: DistilBERT, scikit-learn, Chrome MV3, Flask |
- Software Engineer @ Talendy (Tech Japan) — cross-source dedup over 256K records; leading an Azure → AWS search migration (OpenSearch, ECS Fargate, S3); AI job summaries with Amazon Bedrock.
- Software Engineer Intern @ Akatsuki AI Technologies, Japan — built the automated data pipeline for a healthcare job-search platform (145K+ listings across 47 prefectures) and generated catalog summaries with Azure OpenAI.