Process models built from the lab and plant runs you already have, reinforcement learning for control decisions that repeat many times a day, and AI agents connected to your MES and ERP. Senior engineers, independent since 2013.
The team at Winder.AI are ready to collaborate with you on your AI project. We tailor our AI solutions to meet your unique needs, allowing you to focus on achieving your strategic objectives. Fill out the form below to get started.
AI consulting for manufacturing means finding the production decisions a model can improve, then building the system that makes or recommends them. Winder.AI has been an independent consultancy since 2013. In manufacturing, that has meant reinforcement learning for CMPC’s paper mill in 2022 and, in a 2026 engagement, models of a manufacturer’s process built from a few hundred lab samples.
The decisions differ, and each needs its own method. Choosing a recipe from tens or hundreds of lab runs is a small-data modelling problem, and we solve it with a statistical model of the process and an optimiser on top. Adjusting a continuous process as it runs is a control problem, which reinforcement learning fits when you have a simulator or a long plant history. Triaging maintenance alerts and sourcing parts across your manufacturing execution system (MES) and enterprise resource planning (ERP) system is knowledge work that AI agents can take on.
Which approach fits depends on how much data you hold and how often the decision repeats.
| Approach | Best for | Main weakness | What it needs |
|---|---|---|---|
| Full factorial design of experiments (JMP, Minitab) | A few settings and a lab with spare capacity | The run count multiplies with every setting you add | Every combination, run up front |
| Linear regression or a response surface | A quick first model your engineers can read | Overconfident away from your data, and can predict results the physics rules out | Tens of runs |
| Gaussian process with an optimiser | Recipes and settings with tens to hundreds of runs, where each new run costs money | Only as good as the variation in your past runs: a setting held at one value in every run cannot be optimised | Tens to hundreds of runs, plus your engineers’ knowledge of the process |
| Gradient boosting or a neural network | Plants with years of dense sensor history | Its uncertainty estimates are missing or too narrow, so an optimiser mistakes blind spots for good settings | Hundreds to thousands of well-spread runs, and roughly ten times that for a neural network |
| Reinforcement learning | Control decisions that repeat many times a day | Needs ten to a hundred times the data of a small-data model | A simulator or a long plant history |
| AI agent | Knowledge work across systems, such as maintenance triage and supplier sourcing | It acts on live systems, so high-stakes actions need a person to approve them | Access to your MES, ERP and operational technology (OT) systems |
Each problem below needs a different method. We pick the one your data and your decision can support.
Models of how your settings drive quality and cost, built from tens to hundreds of lab and plant runs, with an optimiser that finds the cheapest recipe that still meets spec.
Control policies for continuous processes, learned from a simulator or a long plant history.
Maintenance triage, production scheduling, supplier sourcing and quality checks, connected to the systems your plant already runs.
Demand, capacity and operational forecasts, and anomaly detection, built with whichever method the data supports, from statistical models to deep learning.
Job scheduling, dispatching and stock levels are chains of decisions, which reinforcement learning optimises. Phil Winder’s articles review the published research.
Purchase orders, bills of materials, quality certificates and engineering specifications, extracted and checked automatically.
Each lab trial or plant run costs money, so the history of a process often runs to tens or hundreds of them, too few for deep learning. We model a process at that size with a Gaussian process, a statistical model that learns from few examples and has your engineers’ knowledge written into it. An optimiser on top finds the cheapest recipe that still meets your specification.
A Gaussian process gives a range with every prediction, like a weather forecast that gives a likely high and low as well as a temperature. The range is narrow where you have run the process and widens where you have not. This range matters more than a small gain in accuracy, because the optimiser uses it to tell a promising setting from a blind spot. A model that gives one number everywhere sends the optimiser towards settings outside your data, where that number is a guess.
The same range chooses the next experiments. A design drawn up in one go grows quickly: in one engagement, a design of experiments we proposed would have taken the client’s lab more than a month to run. The model instead ranks candidate runs by how much they would teach it and how likely they are to beat your current recipe. Your lab runs a small batch, we refit the model, and the loop repeats. This is Bayesian optimisation.
Reinforcement learning learns a control policy: given the state of the plant, which action to take to serve a long-run goal such as throughput or cost. It suits a continuous process where the same decision repeats many times a day and early actions have long-range effects. The price is data, so it needs a simulator of the process or a long plant history to learn from.
Phil Winder wrote the O’Reilly book on reinforcement learning, and has written about its use in job scheduling and assembly and inventory control.
We have built autonomous systems for over a decade, and our first agents in manufacturing were the reinforcement learning agents that suggested settings to operators at CMPC’s paper mill. The language-model agents we build today take on the knowledge work around production: triaging a maintenance alert against the asset’s history, reworking a production schedule around a late delivery, finding another supplier for a part, or checking a batch record against its specification. Each agent connects to your MES, ERP and OT systems, with a human approval step on high-stakes actions.
We build agents on Helix, our own platform for running agents inside an organisation, or on the framework you already run. Every agent ships with evaluation, guardrails and tracing, and can run on your cloud or on-premises, including air-gapped networks, so plant data stays on site. Our AI agent development page covers the engineering, and the legal research assistant we built for Temple University is one we have shipped.
Documents are the narrowest case. Purchase orders, bills of materials, quality certificates and engineering specifications can be extracted and checked automatically, and document automation for manufacturing covers that work in detail.
We start from the data you already have, test away from the plant, and put recommendations in front of your staff before anything runs on its own.
Each step produces something your engineers can check before the next one starts.
We start with what you hold: lab trials, plant runs, instrument history and the systems an agent would read. We measure how much each setting actually varied and tell you, before any modelling starts, which questions your data can answer. You get a baseline to beat: how well the simplest guess predicts data the model has never seen.
Your engineers already know things a model would need hundreds of runs to learn, such as that a reaction speeds up as it warms. In a small-data model we write that knowledge into the model’s starting assumptions, so it holds where data is thin. For reinforcement learning it goes into a simulation of the process, trained on your plant data.
We test every model on experiments it has never seen, holding back a whole set of experiments at a time, and check that its uncertainty ranges hold. A control agent is tested in simulation against start-ups, shutdowns and extremes before it sees the real process.
The first system in front of your staff recommends, and a person decides. An optimiser labels any setting outside what you have run as an unproven experiment, your lab runs a small batch, and we refit.
In our experience, on a few hundred samples the common models predict about equally well, and what separates them is whether you can trust their uncertainty. A 2026 engagement showed both. A manufacturer gave us a few hundred lab samples from a few dozen experiments and asked which settings would make their product more cheaply. The client is confidential. We tested each model on experiments it had never seen.
A Gaussian process, gradient boosting and a simpler regression model each predicted those experiments with about half the error of guessing the average, and the differences between them sat within the noise. The Gaussian process’s uncertainty ranges held on the new data. The other two were overconfident: they claimed more certainty than they had. Pushed beyond the data, the simpler model also predicted results the physics rules out, and the Gaussian process, with the engineers’ knowledge built in, stayed within it. An optimiser trusts whatever the model tells it, which is why we put the Gaussian process underneath it.
The same method works on public data you can check. On 42 published lab brewing mashes, Bayesian optimisation found the best mash in about four tries in a typical run when it started with brewing knowledge, six with none, and twenty when the mashes were tried in random order. Each count includes the three mashes it started from. The working, code and plots are in Phil Winder’s Gaussian process notes.
In 2022 we built a reinforcement learning proof of concept for CMPC’s paper mill, where operators ran the process by hand. We trained a simulation of the process on real plant data, tested agents against start-ups, shutdowns and extremes, and built agents that suggested settings for the injection systems. Together with CMPC we chose recommendations over full control, so operators kept the final decision. The CMPC case study describes the work.
In 2023 we built a similar proof of concept for Genesis Energy’s hydroelectric schemes, on a simulation developed with their engineers. Our Genesis case study concedes that reinforcement learning was probably more than a single scheme needed. Genesis chose it so that the same approach could learn each new scheme they added.
A model learns from variation. If every run used the same recipe, a model has nothing to learn a different recipe from, and the first job is a small set of experiments, which we can design with you. If a rule can make the decision, use the rule: a model is for decisions that are fuzzy. And a process with no simulator and a short history starts with a small-data model, because reinforcement learning has too little to learn from.
Indicative rates and project prices are on our pricing page.
Case study
Winder.AI helped Duetto evaluate offline reinforcement learning for dynamic hotel pricing. Over five months, the engagement progressed from behavioural cloning baselines through Implicit Q-Learning experiments on real booking data, revealing where RL outperforms simpler approaches, what data quality prerequisites exist, and how to evaluate pricing agents when ground truth is unavailable.
Case study
Our financial client is based in the UK. They specialise in providing services to the finance industry. Their data science team embarked on a project to leverage reinforcement learning within their product offering. Winder.AI, world-leading authors and experts on reinforcement learning, helped them deliver their POC into production. Read on to find out more.
Case study
Genesis Energy is a power generation company in New Zealand that sells electricity generated by hydroelectric and hydrothermal generators to the domestic energy market. Currently, people control the decisions surrounding power generation and pricing. Genesis asked Winder.AI to help them develop a reinforcement learning-powered solution to automate generation and pricing.
New Zealand has the enviable situation of possessing high-altitude lakes refilled with ice melt. Discharging the lake presents an ample kinetic energy store that can be utilised for power generation via a turbine. Hydroelectric power generation is therefore sustainable and low carbon.
Case study
Winder.AI helped CMPC, a large paper milling company, to optimise their production process by using reinforcement learning. CMPC are now able to automate industrial processes that were previously manual. This case study describes our approach and the results.
Case study
Ideally, the best way to improve the security of any system is to detect all vulnerabilities and patch them. Unfortunately this is rarely possible due to the extreme complexity of modern systems. One primary threat are payloads arriving from the public internet, with the attacker using them to discover and exploit vulnerabilities. For this reason, web application firewalls (WAF) are introduced to detect suspicious behaviour. These are often rules based and when they detect nefarious activities they significantly reduce the overall damage.
Common questions about AI consulting for manufacturing. For anything else, book a call.
It means finding the production decisions a model can improve, then building the system that makes or recommends them. Winder.AI does four kinds of manufacturing work. Process models find cheaper recipes and settings from tens or hundreds of lab and plant runs. Reinforcement learning handles control decisions that repeat many times a day, as in our 2022 proof of concept for CMPC’s paper mill. AI agents take on knowledge work across plant systems such as the manufacturing execution system (MES) and enterprise resource planning (ERP), with a person approving high-stakes actions. Document automation extracts purchase orders, bills of materials and quality certificates. We also build forecasting and anomaly detection models.
When the same control decision repeats many times a day and you have a simulator or a long plant history to learn from. Reinforcement learning needs ten to a hundred times more data than a small-data model such as a Gaussian process. In 2022 we built a reinforcement learning proof of concept for CMPC’s paper mill, where agents trained on a simulation of the process suggested settings to operators. Phil Winder wrote the O’Reilly book on reinforcement learning. For a batch recipe with tens or hundreds of runs, a small-data model and an optimiser usually come first, and we will tell you which one fits.
AI agents take on knowledge work that spans several plant systems and ends in a judgement: triaging maintenance alerts against asset history, reworking production schedules around late deliveries, finding alternative suppliers, and checking batch records against specifications. We connect them to your manufacturing execution system (MES), enterprise resource planning (ERP) and operational technology (OT) systems, put a human approval step on high-stakes actions, and can run them on-premises or air-gapped so plant data stays on site. Winder.AI builds agents on Helix, its own agent platform, or on the framework a client already runs. Our first agents in manufacturing were the reinforcement learning agents that suggested process settings to operators at CMPC’s paper mill in 2022.
Tens to a few hundred runs in which the settings you control actually vary. In a 2026 engagement, a few hundred lab samples from a few dozen experiments let a model predict experiments it had never seen with about half the error of guessing the average result. How much the settings varied matters more than the number of rows: a setting held at one value in every run cannot be optimised. From there the model picks the next experiments, so your lab runs small batches instead of one large design.
On a few hundred runs they predict about equally well, so the choice comes down to whether you can trust what the model says about its own uncertainty. A Gaussian process gives a range with every prediction, wide where you have little data and narrow where you have plenty. In a 2026 manufacturing engagement, gradient boosting and a simpler regression model claimed more certainty than they had on data they had not seen, and the Gaussian process’s ranges held. An optimiser trusts whatever the model says, so it needs the model whose ranges hold. A neural network needs around ten times more data before it competes.
Our indicative rates and project prices are published on our pricing page at winder.ai/services/pricing/. We scope each manufacturing engagement after a first call, once we have seen what data you hold and which decision you want to improve.
Senior engineers with no sales layer between you and them, including Phil Winder, the founder, who wrote the O’Reilly book on reinforcement learning and built the models in our current manufacturing engagement. Winder.AI has been an independent consultancy since 2013, and clients named on our service pages include Google, Microsoft, Shell, Nestlé and Ofcom. We work with your process engineers and lab staff, because their knowledge of the process goes into the model.
That page covers one job in detail: extracting purchase orders, bills of materials, quality certificates and engineering specifications into your ERP. This page covers manufacturing as a whole: process optimisation, reinforcement learning for process control, AI agents for operations, and document automation as one part of it. The same team delivers all of it.
The team at Winder.AI are ready to collaborate with you on your AI project. We tailor our AI solutions to meet your unique needs, allowing you to focus on achieving your strategic objectives. Tell us about it in the form at the top of this page, or book a call.