Field notes

Production AI engineering, written down

Technical write-ups from building and running LLM systems on a live store — the design decisions, the production failures, and the principles I took from each. Every piece is drawn from real work and keeps the rule I hold to: only claim what's verified.

• ~9 min read · Observability · LLM reliability

  A friendly fallback is a silencer on a fire alarm

  A polite chatbot answered “please call our hotline” to everything — and it was an outage nobody could measure. Three friendly veneers that hide real failures, and the discipline of making each one leave a signal.

  Read article →
• ~9 min read · AI agents · Self-improving systems

  Teaching a production chatbot to learn from its own failures

  Designing a 48-hour self-learning loop so a sales chatbot improves from the answers it got wrong — structured error signals, training knobs split from code, and a permission model safe enough to let an AI agent touch a live system.

  Read article →
• ~8 min read · LLM reliability · E-commerce search

  No card beats a wrong card: debugging an LLM that recommended houses to a customer asking about style

  A customer asks about “indie style” and the bot shows a prefab house. Tracing an intermittent bug down through seven stacked causes, one wrong assumption of my own, and an “anchor + modifier” search model.

  Read article →
• ~8 min read · Function calling · WooCommerce

  Giving an LLM real hands: function-calling on a live WooCommerce store

  How to let an LLM not just answer but act — recommending products and driving a real WooCommerce cart via function-calling: architecture, the structured-output contract, and the production traps (nonces, LiteSpeed cache).

  Read article →

Want the whole picture in one place? Read the full AI Sales Chatbot case study (bản tiếng Việt).