Guillaume
Jerke.

As Plant Planner I am the team’s hub for data, so I first automated the chores: scripts that ingest Excel files on their own, and so on. Every AI idea I had needed the same thing — one source of data, already ingested and ready to use — so I built it: collect, store and distribute, securely, everything planning needs to work.

On top of it, I spent a lot of time teaching the LLM what every table and field means, so it can join them the way a planner would. An MCP server and a Python sandbox let an agent run code safely. The result is an agent that answers tactical and strategic questions in plain language, filtering the right data first.

It impressed my managers because it is easy to use and the data is already there, which is usually what blocks this kind of project.

Using the plant’s historical planning data, I built a predictive application that classifies the quality result of a bagged lot — OK, OK after Quality Decision, or NOK — with 86% accuracy.

It was the first project of its kind at Bayer: machine learning applied to production and planning. We use it to choose input batches more wisely, which cut quality scrap after bagging by about 5%.

Our business environment is complex, and our planning tool even more so. Central Planning sends a continuous stream of requests and constraints — by email, in Excel files, sometimes in a Teams message — and keeping track of them by hand is fragile.

My tool collects every comment, file, text or image, understands it with an AI agent and turns it into structured constraints. Plant constraints are coded so they accept those external ones; the tool then computes the plan and writes it back into the planning system.

A complete rebuild of an application I first created in my first year at Monsanto, as an Access database. HARVS tracks harvest intake at the plant, husking, drying, shelling and storage.

It is connected to our data lake to pull harvest times, so the plant sees the incoming flow before it arrives. I focused on quality: a complete internal traceability tool. Its database is exposed to Excel Power Query and Power BI for advanced KPIs.

I used AI to write the application — and to ship quick fixes during the campaign, while production was running.

With our local Lean manager, I spent several months writing the specification for digital tier meetings (our “DDS”, Tier 1 / Tier 2 / Tier 3). Then I built the application and iterated with my colleague. It was inspired by Fabriq, but simpler and dedicated to our routines.

It covers safety, quality and production escalations, standardised KPI display, action plans, Gemba walks, task transfers between tiers, site/plant/department communications (maintenance work, training reminders…), a virtual board for every meeting, and a large-screen facilitation mode with a simplified interface.

Coded with AI; no LLM inside yet. It runs on site and goes live with pilot departments after the campaign.

Power BI dashboards built entirely with AI, from very precise requests that the standard templates cannot deliver — possible because the data already sits in the Master DB.

Quality tracking files for semi-finished and finished product centralise quality results and update themselves. During the campaign they give an exact waste rate to every stakeholder.