Kewpie cuts factory simulation work 78.8% with Devin

Kewpie's production engineering team has used Siemens' Tecnomatix Plant Simulation to verify production-line and equipment designs before investment. Building those models required dedicated software and a specialized programming language called SimTalk, making the work heavily dependent on a small number of skilled people. When the team set out to use generative AI to raise the level and efficiency of that work, it ran into three obstacles: general-purpose LLMs could not produce accurate SimTalk code; Plant Simulation's GUI meant AI-written code could not run in the environment; and the work was confined to one team, with no path for other members or the wider organization to use it.

To address this, Kewpie adopted Devin, a self-hosted AI agent developed by Cognition AI and supplied with support from DeNA AI Link, a DeNA Group company that helps businesses adopt AI. Rather than starting with the full Plant Simulation environment, DeNA AI Link began by feeding Devin documents through a browser, so that knowledge and know-how accumulated in one place and could be shared across the team. Code execution and verification, which Plant Simulation could not handle, was moved to Devin CLI, where Devin itself can run programs, catch errors and fix them without an engineer's involvement. A set of repeatable "Playbook" procedures for extracting review points was built in as an asset.

The published result — 8.5 person-months falling to 1.8 person-months, a 78.8％ reduction — is a PoC figure for one large-scale model, and Kewpie's production technology department frames it as confirmation that a shift to higher-value work is possible, with confidence that putting AI at the center of dev work points the way forward. DeNA AI Link describes the result as expanding the variations and scale of simulations that can be handled, and says it aims at "making factory simulation second nature," in which engineers in each process build models on their own. The extent to which that larger goal is reached is likely to depend on how well the Playbook and accumulated reference documents travel beyond the conditions of this one test.