AI Knowledge Preservation and Defect Diagnosis System for Japanese Precision Parts Manufacturer
English summary
An independent AI consultant developed a three-tier system for a 400-employee precision parts manufacturer near Osaka to address the impending retirement of two veteran inspectors with 99.7% accuracy. First, 60 hours of inspectors' verbal explanations were recorded and transcribed using LLMs to build a retrievable knowledge base of tacit diagnostic logic. A RAG system enables new inspectors to query similar cases, while a vision AI model classifies defects with root cause suggestions. After five months, A/B testing showed AI-assisted new inspectors reached 99.2% accuracy, up from 96%, and the knowledge base was adopted by production engineers. The project reframed the initial request from a simple AI visual inspection to a knowledge preservation system.
Chinese summary
独立AI顾问为大阪附近一家400人精密零部件制造商构建了三层系统：先录制两位即将退休的资深检验员60小时口述判断逻辑，用大模型转写并构建可检索知识库；再搭建RAG系统，让新人拍照获取相似案例和判断依据；最后训练视觉模型输出缺陷分类与根因建议。项目历时五个月，使新人质检员准确率从96%提升至99.2%，知识库被产线工程师沿用。这比客户最初要求的“AI质检”更聚焦于知识保全，方案核心并非简单缺陷检测，而是将隐性经验文档化。
Key points
Client required AI visual inspection, but real problem was preserving retiring inspectors' tacit diagnostic knowledge.
客户最初要求AI视觉检测，但根本需求是保留退休检验员的隐性诊断知识。
60 hours of inspectors' spoken commentary was recorded, transcribed via LLM, and structured into a searchable knowledge base.
录制检验员60小时口述，用LLM转写并结构化，建成可检索的知识库。
RAG system provides new inspectors with relevant past cases and reasoning, not just pass/fail decisions.
RAG系统为新人提供类似案例和老师傅的判断逻辑，而非仅给出合格/不合格结论。
AI vision model classifies defects with root cause hints (e.g., mold wear vs. material impurity) and links to knowledge base.
视觉模型输出缺陷分类与根因提示（模具磨损/材料杂质等），并关联知识库案例。
Post-deployment A/B test: new inspector accuracy rose from 96% to 99.2%, close to veteran level.
系统上线后A/B测试，新人质检员准确率从96%升至99.2%，接近老师傅水平。
Human factors dominated: inspectors resisted recording until management framed it as teaching successors, not AI replacement.
人因是关键：检验员起初抗拒录音，直到管理层以“向后辈传授技艺”而非“被AI替代”来说服。