AI & ML interests
Predictive maintenance (PdM) for aging mechanical equipment: smart monitoring and control systems that grow under administrator approval. Self QA/QC, on-premise small LLMs, edge neural networks.
Recent Activity
NCDTech: We design and research smart monitoring and control systems that grow.
On top of 24 years of verified measurement and control, we put AI that inspects itself and explains in plain language.
Without internet, under administrator approval.
We focus this technology on predictive maintenance (PdM) for aging mechanical equipment.
(Our website and the diagrams and screens below are in Korean; they are the same figures used on our website and demo video.)
Three things we never compromise
| # | pillar | in one line | the worry it answers |
|---|---|---|---|
| 1 | On-premise | One GPU 8GB PC, zero outbound traffic | "Does our data leave the building?" |
| 2 | Approval gate | AI only proposes; verdicts stay in code. Full revision history, nothing silently deleted | "If AI judges on its own, who is accountable?" |
| 3 | Built-in self-inspection | Wired in at development time, not bolted on later | "Can this attach to the equipment we already run?" |
Our roots are in smart monitoring and control.
With our robot simulation studio we stand on that extension line, and our AI sits on top of it.
The robot you see below is our test bed: the same stack of edge neural network, self QA/QC, and on-premise LLM that we now point at machine health.
Conventional monitoring and control ends at collecting data, showing it, and raising alarms on preset conditions.
Interpretation and judgment are left to the experts who know the system.
What we build is the layer above that: self-inspection wired into every stage, and an on-premise LLM that reads those inspection records and explains them in plain conversation.
Making code verify itself is nothing new.
Test benches and unit tests live outside the code; self-checks live inside it: assertions, Design by Contract, and built-in self-tests (BIST/POST) are standard embedded practice.
We have been putting such self-check devices into code throughout our embedded measurement and control career.
Our technology is the wiring above that standard: the check records feed an on-premise LLM, so the system reports in plain language, holds a conversation, and grows its knowledge under administrator approval.
The system grows through use, but only under administrator approval: code-based verdicts are never overwritten, every approved piece of knowledge keeps its full revision history, and nothing is ever silently deleted.
We do not so much invent new capabilities as put AI, without hesitation, on top of abilities we have already verified, and make it actually run.
Current status: first development phase complete, applied to our in-house autonomous driving simulator.
Demo application in progress, with the knowledge base still growing.
You can watch the whole loop run live in our demo video (22 min, Korean): a GPU 8GB PC, no internet, the system chatting about its own inspection reports and growing under administrator approval.
ํ๊ตญ์ด
NCDTech (์ค์จ๋ํ ): ์ฑ์ฅํ๋ ์ค๋งํธ ๋ชจ๋ํฐ๋ง ์ ์ด ์์คํ ์ ์ค๊ณํ๊ณ ์ฐ๊ตฌํ๋ ํ์ฌ์ ๋๋ค.
24๋ ๊ฒ์ฆ๋ ๊ณ์ธกยท์ ์ด ์์, ์ค์ค๋ก ๊ฒ์ฌํ๊ณ ์ฌ๋ ๋ง๋ก ์ค๋ช ํ๋ AI๋ฅผ ์น์ต๋๋ค.
์ธํฐ๋ท ์์ด, ๊ด๋ฆฌ์ ์น์ธ ์๋์์.
์ด ๊ธฐ์ ์ ๋ ธํ ๊ธฐ๊ณ์ ์ค๋น์ ์์ง๋ณด์ (PdM)์ ์ง์คํฉ๋๋ค.
์ ํฌ๊ฐ ํํํ์ง ์๋ ์ธ ๊ฐ์ง
| # | ๊ธฐ๋ฅ | ํ ์ค | ๋ตํ๋ ๋ถ์ |
|---|---|---|---|
| 1 | ์จํ๋ ๋ฏธ์ค | GPU 8GB PC ํ ๋, ์ธ๋ถ ํต์ ์์ | "์ฐ๋ฆฌ ๋ฐ์ดํฐ๊ฐ ๋ฐ์ผ๋ก ๋๊ฐ๋?" |
| 2 | ์น์ธ ๊ฒ์ดํธ | AI๋ ์ ์๋ง, ํ์ ์ ์ฝ๋. ์ ์ฒด ๊ฐ์ ์ด๋ ฅ ๋ณด์กด, ๋ฌด๋จ ์ญ์ ์์ | "AI๊ฐ ๋ฉ๋๋ก ํ๋จํ๋ฉด ๋๊ฐ ์ฑ ์์ง๋?" |
| 3 | ๋ด์ฅ ์๊ฐ๊ฒ์ฌ | ๋์ค์ ๋ถ์ด๋ ๊ฒ ์๋๋ผ ๊ฐ๋ฐ ๋จ๊ณ์ ์ฌ์ | "๊ธฐ์กด ์ค๋น์ ์ ์ฉ ๊ฐ๋ฅํ๊ฐ?" |
์ ํฌ์ ๋ฟ๋ฆฌ๋ ์ค๋งํธ ๋ชจ๋ํฐ๋ง๊ณผ ์ ์ด์ ๋๋ค.
๋ก๋ด ์๋ฎฌ๋ ์ด์ ์คํ๋์ค๋ก ๊ทธ ํ์ฅ์ ์์ ์๊ณ , AI๋ ๊ทธ ์์ ์ฌ๋ผ๊ฐ๋๋ค.
์๋์ ๋ณด์ด๋ ๋ก๋ด์ ์ ํฌ์ ํ ์คํธ๋ฒ ๋์ ๋๋ค: ์ฃ์ง ์ ๊ฒฝ๋งยท์๊ฐ๊ฒ์ฌยท์จํ๋ ๋ฏธ์ค LLM์ผ๋ก ์ด๋ฃจ์ด์ง ๊ฐ์ ์คํ์ ์ง๊ธ์ ์ค๋น ๊ฑด๊ฐ ์ชฝ์ผ๋ก ํฅํ๊ฒ ํ์ต๋๋ค.
๊ธฐ์กด์ ๋ชจ๋ํฐ๋งยท์ ์ด๋ ๋ฐ์ดํฐ๋ฅผ ๋ชจ์ ๋ณด์ฌ์ฃผ๊ณ ์ ํด์ง ์กฐ๊ฑด์ผ๋ก ๊ฒฝ๋ณด๋ฅผ ๋ด๋ ๋ฐ๊น์ง์ด๊ณ , ํด์๊ณผ ํ๋จ์ ์์คํ ์ ์๋ ์ ๋ฌธ๊ฐ์ ๋ชซ์ผ๋ก ๋จ์ต๋๋ค.
์ ํฌ๊ฐ ๊ตฌํํ๋ ๊ฒ์ ๊ทธ ์ ๋จ๊ณ์ ๋๋ค: ๋ชจ๋ ๋จ๊ณ์ ์ฌ์ ์๊ฐ๊ฒ์ฌ, ๊ทธ๋ฆฌ๊ณ ๊ทธ ๊ฒ์ฌ ๊ธฐ๋ก์ ์ฝ๊ณ ๋ํ๋ก ์ค๋ช ํ๋ ์จํ๋ ๋ฏธ์ค LLM ์ ๋๋ค.
์ฝ๋๊ฐ ์ค์ค๋ก๋ฅผ ๊ฒ์ฆํ๊ฒ ๋ง๋๋ ์ผ์ ์๋ก์ด ๊ฒ์ด ์๋๋๋ค.
ํ ์คํธ ๋ฒค์นยท์ ๋ ํ ์คํธ๋ ์ฝ๋ ๋ฐ์ ์๊ณ , ์๊ฐ๊ฒ์ฌ๋ ์ฝ๋ ์์ ์์ต๋๋ค: assertionยทDesign by ContractยทBIST/POST ๊ฐ์ ์๋ฒ ๋๋ ํ์ค ๊ธฐ์ ์ ๋๋ค.
์ ํฌ๋ ์๋ฒ ๋๋ ๊ณ์ธกยท์ ์ด ์ผ์ ํด์ค๋ ๋ด๋ด ๊ทธ๋ฐ ์๊ธฐ ์ ๊ฒ ์ฅ์น๋ฅผ ์ฝ๋ ์์ ๋ฃ์ด ์์ต๋๋ค.
์ ํฌ ๊ธฐ์ ์ ๊ทธ ํ์ค ์์ ์ฐ๋์ ๋๋ค: ๊ฒ์ฌ ๊ธฐ๋ก์ ์จํ๋ ๋ฏธ์ค LLM์ ์ฐ๊ฒฐํด, ์์คํ ์ด ์ฌ๋ ๋ง๋ก ๋ณด๊ณ ํ๊ณ ๋ํํ๋ฉฐ ๊ด๋ฆฌ์ ์น์ธ ์๋ ์ง์์ด ์ฑ์ฅํ๊ฒ ๋ง๋๋ ๊ฒ์ ๋๋ค.
์์คํ ์ ์ธ์๋ก ์ฑ์ฅํ๋, ๋ฐ๋์ ๊ด๋ฆฌ์ ์น์ธ ์๋์์๋ง ์ฑ์ฅํฉ๋๋ค: ์ฝ๋ ๊ธฐ๋ฐ ํ์ ์ ์ ๋ ๋ฎ์ด์ฐ์ง ์๊ณ , ์น์ธ๋ ์ง์์ ์ ์ฒด ๊ฐ์ ์ด๋ ฅ์ ๋ณด์กดํ๋ฉฐ, ์ด๋ค ๊ธฐ๋ก๋ ์กฐ์ฉํ ์ญ์ ๋์ง ์์ต๋๋ค.
์๋ ๊ธฐ๋ฅ์ ์๋ก ๋ฐ๋ช ํ๋ค๊ธฐ๋ณด๋ค, ๊ฒ์ฆ๋ ๋ฅ๋ ฅ ์์ AI๋ฅผ ๊ฑฐ๋ถ๊ฐ ์์ด ์น์ด ์ค์ ๋ก ๋์์ํค๋ ๊ฒ์ ๋๋ค.
ํ์ฌ ์ํ: 1์ฐจ ๊ฐ๋ฐ ์๋ฃ, ์์ฌ ์์จ์ฃผํ ์๋ฎฌ๋ ์ดํฐ ์ ์ฉ.
๋ฐ๋ชจ ์ ์ฉ ์ค์ด๋ฉฐ ์ง์์ ํ์ฅ ์ค์ ๋๋ค.
Website: https://ncdtech.org
YouTube demo: https://youtu.be/ftsw_vbfw6E
LinkedIn: https://www.linkedin.com/company/112725964
Models: sim-driving-mlp-numpy ยท real-robot-driving-mlp-numpy
Datasets: human-gated-qaqc-knowledge-example ยท real-robot-driving-sessions
Spaces: self-qaqc-llm-demo ยท robot-lab






