Skip to content
View nabeofchanKo's full-sized avatar

Block or report nabeofchanKo

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
nabeofchanKo/README.md

Kosuke Hosoya

製薬・ゲーム・物流の3業界で約7年。現在はNTU(シンガポール)応用AI修士課程で、LLM・RAG・エージェントを使ったシステムを作っています。出力の根拠を人が検証できる形にすることを重視しています。

Bilingual (JP/EN). I build practical AI systems where technical implementation meets domain knowledge — with a focus on making the system's reasoning verifiable, not a black box.

  • 🎓 応用AI修士(NTUシンガポール・2026年修了見込)| M.S. Applied AI @ NTU Singapore (2026)
  • 🧭 医薬品の副作用評価・日英翻訳 / ゲームのデータ分析 / 物流業界での業務効率化・自動化 | ex-Pharmacovigilance · Game Analytics · Process Automation
  • 🌏 日本語ネイティブ・英語ビジネスレベル | Native Japanese / Business English

Selected projects

pv-rag-assistant — 副作用報告の一次評価支援 / Pharmacovigilance triage assistant 症例報告を読み、重篤性(ICH E2A)・既知/未知・因果関係の評価案を根拠付きで作成し、担当者が確認・承認するワークフロー。報告漏れにつながる過小評価を出さないことを最優先に設計し、合成症例の評価で過小評価0件。デモ公開中(登録不要)。 Drafts seriousness, expectedness and causality with the evidence behind each verdict, for a human to approve — 0 under-calls on its gold set. Live demo, no sign-up. LangGraph FastAPI Next.js Chroma Docker AWS App Runner

where-rag-breaks — RAGがどこで壊れるかの検証 / A benchmark for where classical RAG breaks 答えが本文テキスト以外(セルの塗り色・グラフ画像・暗号化ファイルなど11チャネル)にある正解ラベル付き合成コーパスで、古典的RAG・エージェント・ハイブリッドの3方式がどこで壊れるかを同条件で比較。コーパス規模を変えた際の挙動の違いまで測っている。 Measures which information channels break classical RAG, using a synthetic corpus with automatically-derived ground truth — and what an agent fixes, what it does not. Python BM25 + embeddings uv pytest

equity-research-agent — MCPベースのリサーチエージェント / MCP-based research agent 3つのMCPサーバーを連携させて株式リサーチレポートを自動生成。データ取得失敗時は副系統へ自動フォールバックし、出典タグを付けて検証可能な形で出力する。 Orchestrates three MCP servers into an equity-research report, with automatic fallback to a secondary data source and source-tagging so the output stays auditable. MCP (FastMCP) Python Plotly pytest

applied-ai-sms-spam-pytorch — NLPをゼロから実装 / NLP from scratch PyTorchで一から組んだBiLSTMと、ファインチューニングしたDistilBERTを比較。精度だけでなく、誤り方の違いとトレードオフまで検証した。 A BiLSTM baseline built from scratch vs. a fine-tuned DistilBERT, compared on error patterns and trade-offs — not just accuracy. PyTorch HuggingFace


Tech

実務 / Professional: VBA · SQL / BigQuery · Power BI

個人・学術 / Personal & academic: Python · FastAPI · LangChain / LangGraph · OpenAI API · Chroma (RAG) · MCP (FastMCP) · PyTorch · Next.js / TypeScript · Docker · AWS (App Runner) · Git · pytest

Pinned Loading

  1. pv-rag-assistant pv-rag-assistant Public

    副作用報告の一次評価(重篤性・既知/未知・因果関係)を根拠付きで下書きし、担当者が確認・承認するPV支援ツール。合成症例の評価で過小評価0件。 / Pharmacovigilance triage assistant: drafts seriousness (ICH E2A), expectedness and causality for an adverse event report, w…

    Python

  2. where-rag-breaks where-rag-breaks Public

    古典的RAG(chunk+embed+top-k)がどの情報チャネルで壊れるかを、正解ラベル付き合成コーパスで測るベンチマーク / A benchmark measuring which information channels break classical RAG

    Python

  3. equity-research-agent equity-research-agent Public

    銘柄コードから株価・財務・ニュースを集め、8セクションのHTML株式リサーチレポートを自動生成するエージェント(MCP × Claude Code / LLMなしのCLI版あり) | MCP-based equity research agent — three FastMCP servers, Claude Code skills, and a standalone CLI that pr…

    Python

  4. applied-ai-sms-spam-pytorch applied-ai-sms-spam-pytorch Public

    SMS スパム判定で、PyTorch で一から組んだ BiLSTM とファインチューニングした DistilBERT を比較。Transformer が古典的なニューラルネットの完全な上位互換かを、精度・誤り方・モデルの大きさで検証。 / SMS spam classifier comparing a from-scratch BiLSTM against a fine-tuned Dist…

    Jupyter Notebook