Bangkok · October 2026
Enterprise AI Agent architecture AI in production
AI Deputy Director, CJ Express Group. I run the AI Centre of Excellence at head office: a forward-deployed team building computer vision, recommendation and agentic systems for 2,500 stores — plus the platform, evaluation and guardrails that let the rest of the company use them.
We run AI as an enterprise function, not a lab — and the work we put into production is where being wrong costs money. What makes it safe to move fast is unglamorous: a four-layer control model, sanctioned tools, real evaluation, and policy that clears PDPA and ISO/IEC 42001.
Governance isn’t the point. It’s the licence to try things.
Latest — Agentic AI
As AI agents grow more capable with multi-step reasoning and tool use, are we building evaluation systems robust enough to measure their true capabilities—or just documenting short-lived milestones before they outsmart our benchmarks?
Agentic AI
Learn how to build post-training verifiable agentic AI systems, contrasting them with traditional models and outlining the critical training steps.
RAG
Learn how Retrieval-Augmented Generation bridges the critical knowledge gap for Large Language Models.
Agentic AI
Our Documentation Was a Messy Goldmine. Here's How We're Fixing It for AI.
I’m Pisek. I run AI at CJ Express Group, the Thai retail group behind CJ Supermarket and CJ More — 2,500 stores across Thailand. I run the AI Centre of Excellence at head office: a team of forward-deployed engineers embedded in the business, building computer vision, recommendation and agentic systems where the company actually operates.
Before that I did a PhD in electrical engineering at Toronto, on quantum photonics — entangled photon sources in semiconductor waveguides — then spent years designing optical systems for AR glasses and surgical robotics. Physics taught me to distrust any result I can’t measure, which is why I care more about evaluation harnesses than leaderboards.
The governance side exists for one reason: a four-layer model — prevention, gateway, observability, live protection — plus tool sanctioning and policy aligned to PDPA and ISO/IEC 42001. It’s what lets 500 people use what we build directly, and 2,500 colleagues across HQ work with AI, without anyone betting the company on it.
New writing gets posted there first. No newsletter, no RSS feed, nothing to subscribe to — connect or send a message if you want to talk.
