Today more and more machine learning (ML) technologies are applied to business. In ML projects, data scientists train ML models and generate logics by Jupyter notebook in experimental environment, which are intended to be executed only once. Then engineers have difficulty to modify those logics into reusable, composable, and scalable ones in production environment. These hard tasks for engineers can be the bottleneck of the delivery process of the ML models. We modified logics for an experimental environment into pipeline by Kedro – OSS framework for designing ML pipelines – for production environment to make maintainable code with less effort. By Kedro, we could decompose complex logics into set of reusable modules called "nodes" by clarifying the dependencies and relations between Jupyter cells. Also, we could compose new pipeline by replacing “nodes” as the specification changes of system. Besides, we could scale up resources by executing independent "nodes" in parallel. We found that modified pipeline by Kedro enabled us to replace and scale logics more easily. On the other hand, we found that manual conversion from Jupyter to pipeline is still heavy task for data scientists. We will discuss problems of manual conversion, and possible solution for the problems.