Data you already have, used to decide better
Manufacturers generate valuable data every day (orders, bills of materials, production phases, times, costs) that is rarely used to forecast. I have worked with Italian manufacturers, shipping models to production that turn this data into forecasts and decisions.
What I build
- Lead-time and delivery-date forecasting from the bill of materials and production history.
- Cost estimates for an order before it is produced.
- Production planning with reinforcement learning that learns from data instead of hand-written rules.
- Smart configurators that turn customer technical specs into a configuration, a bill of materials and a delivery date.
- Phase forecasting: which phase an order is in and which materials it needs.
- AI assistants inside the ERP that cut cross-team coordination time.
How it works
- Collecting and cleaning dataWe start from ERP, MES and spreadsheet data, often incomplete: half the work is making it reliable.
- The right model for the problemRecurrent neural networks for time series, graph neural networks for bills of materials, reinforcement learning for planning: the model is chosen for the problem, not the hype.
- ERP integrationPython models are exposed through APIs and used directly inside the ERP (for example .NET and Angular), with response times fit for daily use.
- Verification in the fieldForecasts are compared with real outcomes, and the model improves with new data.
Real projects
For Extrema, a graph-neural-network configurator forecasts process time and a delivery date in under ten minutes. For Dekali, an LLM assistant and a reinforcement-learning scheduler built into the ERP cut cross-team coordination from hours to seconds.
DNN + GNN turn customer specs into a complete bill of materials and an accurate delivery quote in minutes. LLM chatbot cuts internal response times by 60%.
LLM assistant + RL scheduler embedded in a manufacturing ERP. Cuts cross-team coordination from hours to seconds.
CAN-bus reader & automated battery test certification, written for Voltfang's incoming-product test pipeline.
Technology
Further reading
A retrospective on replacing a rule-based scheduler at Dekali with an RL policy — and the silent rules nobody had written down.
Predicting which phase a job is in, and what it will cost by the time it ships, from telemetry no one had ever bothered to keep.
How we encoded stairlift components as a graph and used message-passing to predict process time and delivery date in under ten minutes.
A deep neural net that turns customer technical specs into the optimal stairlift configuration, the full Bill of Materials and a delivery date the sales team can actually commit to.
Frequently asked questions
Do we need a lot of data to start?
You need consistent historical data more than huge volumes. The first phase of the project checks exactly the quality and quantity of the data available, before choosing a model.
Will it work with our ERP?
Models are exposed through APIs and can be integrated into most business systems. In the projects I have delivered, integration was into .NET and Angular ERPs.
Is reinforcement learning realistic for an SME?
It can be when planning rules are many, change often and are partly unwritten. In those cases a model that learns from historical data can outperform hand-built heuristics.
How reliable are the forecasts?
It is measured on the company’s real data, comparing forecasts and outcomes over time. Forecasts come with their uncertainty, so the engineering team knows when to trust them and when to check.