AI/ML Engineer focused on RAG systems and LLM applications.
Currently building retrieval‑centric, evaluation‑driven LLM products.

Tesla-RAG – Q&A system over Tesla IR reports, improved from ~10% → ~80% answer accuracy through 8 iterative versions.

• V1: Vector search baseline → 10% (1/10)
• V2: + BM25 hybrid search (RRF) → 30% (3/10)
• V3: + Table-aware chunking → 60% (6/10)
• V4: + Cross-encoder reranking → 80% (8/10)
• V5–V8: + CRAG pipeline, FastAPI API, pytest, Dockerization

Key insight: The largest gain (+30 pp) came from fixing data quality and chunking strategy, not from more complex algorithms.

Stack: Python · ChromaDB · BM25 · SentenceTransformers · Cross-Encoder · CRAG · FastAPI · pytest · Docker

• MCP (Model Context Protocol) – Client/Server architecture for AI tool integration
• AI Agents – Tool use, multi-agent orchestration, agentic RAG
• LLM Application Development – Claude API, prompt engineering, evaluation

• AI/ML: Python · RAG · LangChain-free architectures · SentenceTransformers · Cross-Encoders
• Backend: FastAPI · Docker · pytest
• Past experience: JavaScript · React · Vue · Node.js · PostgreSQL · Swift

RAG・LLMアプリケーションに特化したAI/MLエンジニアとして、評価可能なRAG・エージェント基盤の開発に取り組んでいます。