NCDTECH · SELF QA/QC + ON-PREMISE LLM
A machine that inspects itself, and talks about it. Offline.
스스로 점검하고, 대화로 설명하고, 관리자 승인 아래에서만 자라는 시스템의 실동 데모입니다.
저희는 이 기술을 노후 기계식 설비의 예지보전(PdM)에 집중하고 있습니다.
GPU 8GB PC
no internet
local LLM (4B)
verdicts immutable
admin approval gate
Full demo, 22 min, Korean. The system chats about its own inspection reports and grows its knowledge under administrator approval.
01What is stacked
Our roots are smart monitoring and control. An edge neural network, self QA/QC reports, and a conversational LLM layer sit on top. We now focus this stack on predictive maintenance (PdM) for aging mechanical equipment. (Diagrams are in Korean; they are the same figures used on our website and in the video.)
02The conversation becomes knowledge
An operator asks why the robot drives in short segments. The local LLM answers citing the stage report, and the conversation is captured as a pending knowledge record.
03Nothing grows without approval
The administrator reads the pending record in full, then accepts it into official knowledge, rejects it, or revises it as a new revision. Code-based verdicts are never overwritten.
04The model this produces
A 270-parameter MLP driving in the 3D evaluation stage, with its live inference panel: four ultrasonic inputs, softmax over five commands, and a turning-angle head.
05And the real robot
The tracked test robot driving on the test floor, monitored live by our control dashboard.
Three rules the whole system keeps:
Verdict immutability. The LLM explains inspection results; it can never change them.
Approval gate. Nothing enters the LLM context until an administrator accepts it.
Append-only history. Every record keeps its full revision history. There is no delete.
📚 Dataset (record formats): NCDTech/human-gated-qaqc-knowledge-example
🤖 Model (driving MLP): NCDTech/sim-driving-mlp-numpy
🏢 Organization: huggingface.co/NCDTech · Website: ncdtech.org
NCDTech (New Cyclorama Dream Technology) · smart monitoring and control systems that grow