Notes on industrial AI, edge intelligence, and trustworthy systems.
Research writing from the Director of the AI Data & Security Research Center at KETI — foundation models for industry, agentic-AI security, and the data infrastructure underneath.
At UAI 2026: My Pipeline Was Far Too Sure of Itself
Sitting through a conference on uncertainty and causality, I realized the real problem with the LLM and agent pipelines I've been building isn't that they get things wrong. It's that they have no way to say they don't know.
All writing
12 posts- Jul 23, 2026 ~8 minWhite Paper: Advancing AI Safety and Security
A technical white paper on AI safety and security in the agentic era — mechanistic interpretability, automated red-teaming, dynamic evaluation, agentic action-restriction, enterprise confidentiality, and the governance landscape across the EU, the US, and Korea. Free to download (CC BY-NC 4.0).
- Jul 21, 2026 ~11 minThree Failures Before a Knowledge Base That Held
I wanted a knowledge base over the documents my research projects generate — proposals, agreements, reports, trip records. It took three abandoned designs to learn that the hard part was never extraction. It was change, cost, and knowing what you don't know.
- Jul 13, 2026 ~12 minRunning AI Inside a Trusted Execution Environment
Less a tutorial, more a field report — why you'd run an AI model inside a Trusted Execution Environment (to keep weights and data out of the host's reach), why today's TEEs strain under it (tiny enclave memory, CPU-only trust, costly CPU↔GPU transfers), and recent research directions, including why confidential GPU inference needs a Hopper-class data-center GPU and why Jetson Thor's Blackwell doesn't qualify.
- Jul 01, 2026 ~11 minLeaking Through an Authorized Door — the Security Problem of Generative and Agentic AI
Industrial foundation models and agentic AI can transform process management, but they punch a different kind of hole in information protection. This is about information that leaks through a legitimate access route, the attacks worth thinking about, the standards and guidelines that speak to them, and a defensive paradigm that brings an AI point of view into cybersecurity.
- Jun 29, 2026 ~12 minAn Industrial Data Lake for Industrial AI
A sketch of an Industrial Data Lake — a way to build industrial AI while keeping data and models protected as corporate assets. It covers the conflicting requirements industry faces, agentic AI as an answer, a stakeholder-and-business-model structure, what Germany and Europe's IPCEI-AI suggests, and the security functions the system has to carry.
- Jun 06, 2026 ~9 minIndustrial AI and its visibility
Commercial AI is already astonishingly good — so do we even need a separate industrial AI model? A look at three constraints (protecting information, cost, and sustainability) and why the thing that ultimately holds it all together is AI visibility.
- Jun 03, 2026 ~6 minWhat an Industrial Foundation Model Should Be
The phrase 'industrial foundation model' turns up in every R&D program now. But to earn the word 'foundation,' such a model shouldn't be an omniscient know-it-all — it should be the bedrock that domain knowledge gets built on.
- May 29, 2026 ~19 minWhen the Knowledge Base Wants to Be a Graph
Two days after building a markdown knowledge base, the cracks started showing. Notes on turning it into an RDF graph using Apache Jena Fuseki - the architectural calls, the model comparisons across Claude, qwen2.5, and exaone3.5, and why the documents themselves are the only sustainable source of truth.
- May 27, 2026 ~7 minBootstrapping a Personal Knowledge Base in an Afternoon
Notes from a day spent designing folders, slugs, and a small LLM skill so that future updates to a personal knowledge base only require pointing at the source material.
- May 23, 2026 ~5 minA Study Roadmap for Uncertainty Quantification + Inverse Dynamics
A leveled reading list and a 10-week curriculum for getting from Bayesian basics to physics-informed, uncertainty-aware inverse dynamic models.
- May 23, 2026 ~8 minUncertainty Quantification Meets Inverse Dynamics
A concept review of uncertainty quantification (UQ), inverse dynamic models (IDM), and why combining them matters for safe, data-efficient robot control.