A PoC on quality factor analysis and AutoML

CoreData EngineeringData Platform EngineeringData AnalysisBI & Data VisualizationMachine Learning PoCVendor Management

Organisation: YAMATO KOGYO Co., Ltd.
Period: Sep 2022 – Dec 2023
How it started: Assigned (The decision to run it came from the steelmaking section manager)
My part: Framing the theme, preparing the data, negotiating with the vendor, and evaluating the results
For whom: The manufacturing division (steelmaking) and the external vendor
Scale: 6,545 steelmaking-related defects (August 2022 to March 2023) and equipment signal time series at one-second and 100 ms intervals

What I did

Prepared the datasets from the plant database and the equipment time series, and handed them to the vendor with ER diagrams.
Kept the approach and the deliverables aligned with the vendor through meetings, minutes and requests for further analysis.
Wrote a theme definition sheet — the problem, the prediction target, the data, the conditions for adoption, the criteria for taking it on — so both sides shared the same premises.
Estimated the business impact.
Went as far as correlation analysis for candidate factors and defect-prediction models per mould size, with accuracy evaluated. Predicting the absence of defects works well, predicting their occurrence does not yet, and accuracy improves the more narrowly the target is scoped.

Split it into stages — look at the whole with shared factors, narrow to the mould sizes with the most defects, judge whether a predictive model is feasible — so it could be stopped cleanly at any of them.

Extraction that goes inside the implementation, visualisation that serves the business goal.
I work alongside the business from the definition of the metric itself, and stay with it through to the insight it was asked for. From IoT systems to web and mobile, I have broad experience of building software, so I extract data with a deep understanding of the implementation behind it. Visualisation and analysis are the same — I put generative AI to work and deliver quickly, in the form the person will actually use. Underneath that, I can shape the data, the context around it and the ETL that feeds it for the age of generative AI, so that people and AI alike can work with what is there.

I go and learn the business itself
An IoT system for a screw factory as a student, and a secondment to a steelmaking division in my first year of work, taught me the same lesson: what matters is the business outcome, the data that comes out of it, and the domain knowledge you need to connect the two. Sometimes the right answer is to change the workflow rather than to build anything at all. I would rather understand what the person asking is trying to achieve — including where AI genuinely helps — than simply implement what I was handed.