Case · Manufacturing
Semiconductor — defect rate 2.40% → 1.20%, reverse-solved
// PERSONA
Han, Plant Manager — semiconductor SME
// INDUSTRY
Manufacturing / Semiconductor
// DATASET
case_01 · semiconductor_defect_rate.csv · 12,480 rows
PROBLEM
The Problem
Line average defect rate exceeded the quarterly target (1.5%) by 0.9pp. Uploading six months of process logs to ChatGPT was a policy violation, and existing BI tools (Tableau) only showed regression — never "which variable to adjust by how much" to hit the target.
// PROCESS LINE
DEFECT 2.40%
// CHATGPT
"6개월 로그 업로드는 사내 보안 위반"
APPROACH
XimTier Approach
Loaded data on-prem, WhatDataAI identified the five core variables, and Reverse What-If reverse-solved each variable's optimum to hit 1.20%. Every result ships with SHAP-based mathematical justification, ready for EU AI Act compliance.
// TARGET
1.20%
OPTIMAL INPUTS
온도
압력
속도
습도
등급
VARIABLES
5 Process Variables Reverse-Solved
| VARIABLE | BASELINE | OPTIMAL | Δ |
|---|---|---|---|
| Temperature | 215.0°C | 224.9°C | +9.9°C |
| Pressure | 88.0 MPa | 83.7 MPa | -4.3 MPa |
| Line speed | 14.0 m/min | 8.0 m/min | -6.0 m/min |
| Humidity | 52.0% | 49.5% | -2.5% |
| Material | 3.0 | 4.5 | +1.5 |
OUTCOMES
Outcomes
OUT / 01
Defect rate −50% (baseline 2.40% → 1.20%)
OUT / 02
Inference 0.18s (real-time slider response)
OUT / 03
Predicted accuracy 92.4% / R² 0.887
OUT / 04
Zero external data leaks (100% on-prem)
⚠ Example data — fine-tuned on customer data at deployment.