# Levi Lee (Wei Lee) — AI / LLM Engineer

> AI / LLM engineer and software-startup co-founder, specializing in LLM agent systems, RAG pipelines, and deep-learning applications. Builds and ships multi-agent orchestration, retrieval systems, and the LLMOps and cloud infrastructure behind them — from problem definition through production deployment. Based in Taipei, Taiwan. Personal site: https://tensorowl.dev

Levi Lee (full name Wei Lee; also goes by "tensorOwl") is a software engineer focused on retrieval and agent systems that survive contact with production. Political-science background (BA, National Taiwan University), self-taught into software, now shipping LLM applications and the LLMOps around them. He is also co-founding a software startup (Faca-iye Technology) in the AI / LLM space. He works across multi-agent orchestration, modular workflow design, and end-to-end ML system development, and leans on LLM-powered tools (Claude, GitHub Copilot) to move fast. The handle "tensorOwl" — a tensor-flavoured owl — began as a TensorFlow typo he decided to keep. Open to coffee chats.

## Now / Current roles

- **Co-Founder & Engineer — Faca-iye Technology** (Taipei, Apr 2026 – present): Co-founded a consultancy specializing in digital transformation and system integration (SI). Leads technical consulting and AI implementation, and designs technical roadmaps for enterprise clients adopting AI and cloud solutions.
- **Software Engineer** (Taipei, Sept 2025 – present): Develops deep-learning and LLM-based agentic AI systems for finance and intelligent manufacturing, owning the cycle from research and prototyping through production deployment.
- **Instructor & Teaching Assistant — NTU AI Club** (Taipei, Mar 2026 – present): Co-designed and delivered an introductory curriculum for newcomers (environment setup, Python fundamentals, the uv toolchain, API-key security, AI-assisted development). Designed and led a LangGraph workshop series on stateful agent orchestration and multi-agent workflow design; also taught AI agent deployment (OpenClaw on Zeabur) and LLM application development. Built ArcanaVision for the club's annual exhibition: a multimodal tarot app using Gemini 2.0 Flash and Vertex AI Imagen 3, deployed on GCP Cloud Run with on-site QR-code access.

## Background

- **Research Assistant — Academia Sinica** (Taipei, Aug 2024 – Aug 2025): Designed and implemented BERT-based NLP pipelines for extracting latent economic signals from judicial decisions. Applied Synthetic Difference-in-Differences (SDID) for causal inference with custom evaluation frameworks. Published the open-source `sdid` package on PyPI (staggered treatment adoption and event-study analysis). Built a real-time carbon-trading experiment platform on oTree with WebSocket live price updates and an order-matching engine.
- **Education**: National Taiwan University — BA in Political Science and Public Administration (Aug 2019 – Aug 2024).

## Selected projects

- [Social-Debate-AI](https://github.com/AluminumShark/Social-Debate-AI): Multi-agent LLM debate system (Support/Oppose/Neutral) on LangGraph, built as an instrument to test — not assume — whether RAG, a GNN, and RL improve persuasion. Stack: provider-agnostic LLMs (local Ollama by default, OpenAI-compatible, or BYOK), FAISS RAG (embeddinggemma, 768-d) over the Cornell ChangeMyView corpus, GraphSAGE/GAT graph networks, and a PPO policy over four strategies; Flask + SQLite with SSE token streaming. The honest ablation is the real subject: RAG showed a small directional gain (within judge noise), the GNN collapsed to the majority class under class imbalance, and the RL reward — derived from that GNN — optimized a mis-measured target. A study in measurement and diagnosis, not "three cool modules."
- [Carbon Trading Experiment](https://github.com/AluminumShark/Carbon-Emissions-Trading-Experiment): Real-time experimental-economics platform on oTree for studying carbon policy (carbon tax, cap-and-trade, control) — a live WebSocket permit market backed by an order-matching engine, YAML-configurable treatments, and full data export.
- [Synthetic Difference-in-Differences (sdid)](https://github.com/AluminumShark/Synthetic_Difference_in_Difference): Causal-inference library published on PyPI implementing the Arkhangelsky et al. (2021) estimator — automatic unit/time weighting, bootstrap standard errors, staggered adoption, and event-study analysis.
- [Build-My-Own-Redis](https://github.com/AluminumShark/Build-My-Own-Redis): A Redis-like in-memory key-value server built from scratch in C++ — TCP socket communication, a poll()-based non-blocking event loop with per-connection state machines, and length-prefixed binary protocol framing (GET/SET/DEL + request pipelining).
- [FinFlash](https://github.com/AluminumShark/FinFlash): Multi-agent financial-news analysis that issues BUY/HOLD/SELL verdicts via parallel agents (sentiment, entity extraction, risk, summarization); FastAPI + LangGraph + LiteLLM backend with a React/Vite frontend, multi-provider LLMs (OpenAI, Anthropic, Gemini, DeepSeek, local Ollama), PostgreSQL + pgvector RAG memory, Google News RSS (Exa optional), Pydantic-structured outputs, and LangSmith tracing.
- [Tick-Backtest](https://github.com/AluminumShark/Tick-Backtest): Tick-level XAUUSD backtester — resamples to 1H OHLC, MA-crossover with an RSI filter, genetic-algorithm optimized and out-of-sample validated.

## Writing

- [Writing index](https://tensorowl.dev/writing): Notes on what survives contact with production, including beginner guides on Python for AI and setting up an AI dev environment.

## Skills

- **LLM & AI**: LangChain, LangGraph, RAG (FAISS, ChromaDB, pgvector, Vertex AI Search), multi-agent orchestration, prompt engineering, LLM evaluation.
- **ML & Data**: PyTorch, HuggingFace Transformers, PyTorch Geometric, scikit-learn, Numba, Optuna, reinforcement learning (PPO).
- **Languages**: Python, C++, R, JavaScript, TypeScript, SQL.
- **Infrastructure**: GCP (Cloud Run, Vertex AI, Cloud SQL), Docker, FastAPI, PostgreSQL, SQLAlchemy, Alembic, WebSocket, Git.
- **Prototyping & tools**: Streamlit, GitHub Copilot, Claude, ChatGPT.

## Site pages

- [Home](https://tensorowl.dev/)
- [Work](https://tensorowl.dev/work)
- [Writing](https://tensorowl.dev/writing)
- [About](https://tensorowl.dev/about)

## Contact

- Email: levi@tensorowl.dev
- GitHub: https://github.com/AluminumShark
- LinkedIn: https://www.linkedin.com/in/wei-lee-3b2881193/
- Threads: https://www.threads.com/@tensor0wl
- Instagram: https://www.instagram.com/tensor0wl