Making complexity manageable.

ORSTEM automates business processes, integrates AI agents into existing system landscapes and applies mathematical optimization models to make robust decisions, even when data is incomplete.

Better decisions in three steps.

From the first data analysis to a productive, integrated system.

Understand processes, assess potential, clean data

We analyze existing workflows, identify automation potential and clean up the data foundation. Where are the data gaps, where is quality sufficient, and which processes are actually worth automating?

Build AI and OR Decision Models

Based on clean data, AI and Operations Research models are developed that explicitly account for uncertainty. The solution is built production-ready and integrated into existing system landscapes. APIs, data flows and handoff points included.

Scaling and Continuous Improvement

The production solution is transferred to live operations, continuously monitored and evolved with new requirements, so the value does not end at go-live but keeps growing.

Plan for uncertainty instead of ignoring it.

In practice, data is missing, forecasts are off, systems deliver contradictory signals. This is not an edge case; it is the normal case.

Classical models assume complete, consistent data. In operational processes this is rarely the case: sensors deliver outliers, historical data reflects bottlenecks rather than actual demand, and different systems report the same situation differently. Ignoring this means optimizing past the problem.

Tackle the challenge together
250200150100500DemandTimePASTFORECAST

Selected learnings:

Can we actually trust the data?

Not always. A classic example is the confusion of sales data and demand data. Companies see what was actually sold, not what could have been sold. If a product sells 50 times in a day but stock was capped at 50, the real demand stays invisible. Models trained on such censored data systematically learn too little and optimize against a distorted picture of reality.

Does a better forecast automatically lead to a better decision?

Not necessarily. The classic approach is predict-then-optimize: forecast first, then optimize based on the forecast. The problem: a model that minimizes forecast error solves a different problem from one that maximizes decision quality. Predict-and-optimize goes one step further and trains forecasting and optimization jointly, directly towards decision quality. This approach is also called Decision-Focused Learning. Forecast accuracy and decision quality are related, but not the same thing.

Do I always need a mathematically optimal solution?

In practice, a good heuristic is often sufficient: fast, robust and understandable to specialists. What matters is knowing how well it actually performs and how close it is to the theoretically achievable optimum. That is exactly what comparison with exact methods provides: not to always solve optimally, but to assess and systematically improve the quality of the heuristic.

Why optimal decisions can fail in reality.

The mathematically optimal result is often the most fragile.

Classical optimization models search for the best point within a feasible region and almost always find it at its boundary. Mathematically this is called an extreme point: the solution sits exactly on the edge of what is possible, with no buffer whatsoever.

The problem: real systems deviate. Demand fluctuates, lead times shift, machines break down. As soon as even one parameter changes slightly, the original solution leaves the feasible region and becomes infeasible. The model then produces no answer, even though the system keeps running and decisions are needed.

Robust and stochastic optimization address exactly this: they model uncertainty explicitly and find solutions that remain feasible and good enough even under deviations. Not perfect under ideal conditions, but reliable under real ones.

Limit ALimit BCap. ACap. BFeasible regionInfeasibleRobust solutionwith bufferOptimumExtreme pointSmall deviationinfeasible

The optimal extreme point sits on both capacity limits simultaneously, with no buffer. A robust solution lies further inside and remains feasible under deviations.

Check robustness in your process
Timur Wohlleber, founder of ORSTEM

Founder

Timur Wohlleber

“I founded ORSTEM because the methods that work in research often never make it into practice. I wanted to change that.”

Applied Data Scientist with a background in mechanical engineering, robotics and AI. Currently pursuing a PhD at TU Munich in cooperation with AUDI AG on robust production and supply chain systems, the scientific foundation for what ORSTEM builds and delivers today.

More about ORSTEM

Just one message away.

Write to us and we will get back to you quickly and without complications.