Every year we set our roadmap aside, lock ourselves in a room for a few days and build during our annual Innovation Hackathon. With one goal in mind: solving real supply chain challenges through data science. This year we raised the stakes: we invited four customers to each bring a real, unsolved problem and hack on it side-by-side with our AI & Data Science team over three days.
Here’s the cases the team dived into and what they managed to accomplish at the end of the hackathon.
Case 1: An AI agent for interacting with safety stock outcome
Even with a strong inventory optimization engine and dashboards in place, at Grünenthal one question kept coming back from their planners: “why does the model want a safety stock of X?” Today, answering that question means either an expert explaining it by hand, a scalability bottleneck, or a manual deep-dive across multiple dashboard pages.
So, during the hackathon the team tasked with this case built an assistant agent (on Databricks Genie) that a planner can simply talk to. It explains the drivers behind each recommendation, highlights the biggest inventory-reduction opportunities first, and runs instant what-if analysis, like for example: “What happens to my safety stock if I move service level from 99% to 99.3%?” Something that previously required a full rerun of the optimization. It also attaches a confidence level to every recommendation using a data-quality alert system, so planners know at a glance which suggestions to trust and which to review.
One moment particularly stood out. When an expert insisted the agent add 10 units to every material “because I said so,” the agent politely refused to override the underlying calculation and asked why instead, exactly the behavior you want from a decision-support tool. It points to a clear vision: keep the familiar, trusted models underneath, and put an agentic workflow on top that makes setup, experimentation and adoption dramatically faster.
Want to learn more? Dina Smirnov is ready to answer all your questions regarding this case.
Case 2: Enhancing demand sensing with open sales orders
Lekkerland’s weekly demand forecast drives warehouse resources and picking plans. It handles normal patterns well, but occasional large demand peaks are hard to anticipate from history alone. The team’s hypothesis: some of that demand is already visible in open sales orders before each forecast run, so could feeding those orders into the model sharpen it?
First, they uncovered just how late demand actually lands. Only 34% of a delivery’s final demand is on the books 7 days ahead, rising to 70% two days ahead, and a full 30% isn’t booked until the day before delivery. On top of that, roughly 77% of volume falls into just two recurring “visibility profiles,” meaning customers tend to order in predictable timing patterns even when they order late.
So, did adding open sales orders help? On their own, they lifted forecast accuracy by only ~0.02%, which is not meaningful. But the reason why is where the value lies, the hackathon team found out. With just 26 weeks of order history, and a model trained on a single location because of compute limits, there simply wasn’t enough data to capture how individual customers behave. The signal is clearly there: many customers order late but consistently, it just needs to be modelled properly.
Want to learn more? George Drakos is ready to answer all your questions regarding this case.
Case 3: Spotting hidden promotions with Dynamic Time Warping
At one of our clients, missing promotional records leave forecasting models blind during the highest-impact periods, driving overstocking, stockouts and lost margin. The team’s idea was to close that gap with Dynamic Time Warping (DTW), a shape-matching technique that automatically detects hidden promotions in historical demand, with no promo calendar required.
The pipeline learns a “promotion fingerprint” by clustering historical demand around known promotions, then scans the full sales history for segments that match that shape. Those detected promo signals feed a driver-based LightGBM forecast, working alongside seasonality, calendar events and the DTW clusters themselves.
The results lined up closely with the client’s own analytics team, which is always reassuring, and put a number on the problem: unplanned, unknown promotions were worth roughly 5% of forecast accuracy. The recommendations that followed were practical and immediately useful: align promotions more closely with sales, give the model a better grasp of the calendar, forecast per warehouse location before aggregating (seasonality and holidays are often local), and bring in price data to weigh item importance.
Want to learn more? Wout Olde Hampsink is ready to answer all your questions regarding this case.
Case 4: Explainable supply chain optimization
The hackathon team explored how agentic AI could make a complex, in-house-built supply-network optimization model easier for planners to interact with, letting people ask questions, run scenarios, and adjust constraints in natural language instead of needing deep knowledge of the underlying optimization model.
By the end of the hackathon, the team had created a demo-ready prototype: a multi-agent system (an orchestrator, a data agent, and an optimization agent) running on Databricks that could solve and compare scenarios, store scenarios with context, provide a chat interface with status updates, and visualize scenario comparisons via graphs.
Want to learn more? Ragnar Eggertsson is ready to answer all your questions regarding this case.
What’s next?
With the hackathon coming to an end, all case outcomes are ready to be implemented by our clients, used, and further improved along the way. We are more than proud to see how far our team has come in just over 3 days and are looking forward to seeing these solutions in action, delivering value in real supply chain planning processes.
In case any of these cases spark your interest: we are happy to tell you more about how each solution was designed and to help you assess whether they could be of value in your processes too. Get in touch with the case owner through the details below each case, or contact our team here.