Hi, I'm Weimeng

Applied AI Engineer building production LLM systems — from domain research through deployment. 🌐 EudaimonAI.org

---

🛠 Tech Stack

• Core Languages — Python · TypeScript · SQL
• LLM Systems — OpenAI · Gemini · Groq · LangChain · RAG with pgvector · Prompt Engineering · LLM-as-Judge Evaluation · Multi-Model Orchestration
• Backend & Infra — Node.js · FastAPI · Next.js · Supabase (Postgres + pgvector + Auth) · Vercel
• Data & Modeling Foundations — Pandas · NumPy · Scikit-learn · XGBoost · PostgreSQL · Feature Engineering · Time Series
• Tools — Git · Cursor · Claude Code

---

🧠 How I Build AI Systems

I approach AI development as a system design problem, not a model problem.

1. Problem framing first. I start by identifying decision points under uncertainty — where users struggle with interpretation, ambiguity, or conflicting signals. The goal is to improve human judgment and clarity, not to maximize raw model output.

2. Systems over models. Instead of relying on a single LLM call, I design multi-step systems:

• Structured input and context capture
• Retrieval (RAG) for grounding
• LLM reasoning with versioned, eval-tested prompts
• Structured outputs with schema validation and typed fallbacks
• Multi-model fallback routing for reliability

3. Eval-first development. I built an LLM-as-judge regression harness with domain-informed synthetic personas before the product had a marketing page. Every prompt change is validated against the eval suite before deploy.

4. Production iteration loops. Real metrics (MAU, retention, session depth) drive the roadmap. Prompts get rewritten based on user signals, not opinions. Ships weekly against real data.

📌 Selected Work

Eudaimon — Production LLM application, live in 18+ countries Solo-built AI system combining RAG with pgvector, multi-model fallback (OpenAI → six Gemini variants → Groq), rolling long-term memory compression, LLM-as-judge eval harness with synthetic personas, and bilingual (EN/ZH) safety-signal routing. Full stack: Next.js · TypeScript · Supabase · Vercel.

Core code is in a private repo; happy to walk through architecture in conversation.