AI/LLM Engineer | Data Engineer | Building production agent systems with LangGraph, custom memory, and eval frameworks
I studied psychology at NTU. Ended up doing data work after graduation and kept going for eight years, from analyst dashboards to building entire products.
At Novo Nordisk I rebuilt reporting on Snowflake and dbt. What used to take a week takes a day. At Appier I built anti-fraud detection that caught over $1M in fake ad spend, and wrote bidding algorithms that improved campaign ROI between 5% and 40%.
I also build my own products, mostly outside work hours. Three are live: AI Commerce Ready (Shopify feed optimization for ChatGPT/Gemini shopping), NexRex (AI running coach with LangGraph agent, 3-tier memory, custom eval suite), and Replai (no-code chatbot builder with RAG).
The psych degree helps more than I expected. How people read information, what they trust, where they get confused. I use that when I design prompts and data products.
I've been using AI coding agents full-time for over a year. Eight months on Claude Code Max, before that Cursor. Not just prompting, but building the system around it: 15 custom commands, automated formatting hooks, a memory system that persists context across sessions, and a feedback loop where the agent plans, builds, tests, and reviews its own work in cycles. I open-sourced the whole setup.
8 months on Claude Code Max ($200/mo), 6 months on Cursor before that. I don't just use AI tools. I build the harness around them.
My Claude Code setup has 3 layers of permission control, 6 automated hooks (every file edit runs ruff + prettier), and 15 custom slash commands. The /autopilot command reads a memory file to discover work, implements it, verifies with tests, and commits. /cycle runs that in a loop with automated QA and design review. I can step away and come back to a finished feature.
The whole config is public: github.com/flsteven87/claude-code-setup
This isn't vibe coding. It's what Karpathy calls agentic engineering: structured human oversight over autonomous agent workflows. Plan, execute, verify, with governance at every layer.