What We Build: Machine Learning Projects and Their Impact

Key points: Machine learning only delivers value once a model actually runs in production and lands in your processes. This overview shows, anonymised, what we build, for each project the technique used and the impact in numbers. From a forecast correction that led to roughly €1.2 million in annual savings at a global manufacturer, to recommendation engines, document intelligence and algorithmic optimisation. The throughline: the model is not the end product, its integration is.

Forecasting & supply chain

Forecast-bias correction at a global tool manufacturer. Statistical demand forecasts often show a systematic over- or under-estimation. We built a model that predicts the direction and magnitude of that bias and corrects it, complemented by analysis of phase-out timing and cannibalisation between products. Impact: an order of magnitude of €1.2 million in annual savings through more accurate planning. Technique: bias modelling on top of the existing statistical forecast.

Demand forecasting across four warehouses at a multi-market retailer. For an e-commerce party with warehouses in four countries, we built per-market SKU demand forecasting with reorder points, including seasonal and promotional effects. Impact: less over- and understock and so lower inventory costs. Technique: time-series models with external regressors, calibrated per market.

Energy forecasting on visitor data at a theme park. The net energy consumption of a large leisure park turned out to be strongly driven by visitor numbers, something a standard load forecaster (OpenSTEF) did not capture well. We analysed why and built an improvement strategy that incorporates visitor forecasts. Technique: domain-specific feature engineering on top of time-series forecasting.

Recommendation & personalisation

Cross-sell and quote scoring at an automotive marketplace. For a dealer in exclusive car parts, we analysed 1.2 million product-vehicle relations for a cross-sell engine, plus a quote-scoring model at 84% accuracy that shows sales upfront which quotes are likely to convert. All directly in the existing backend. Technique: gradient boosting + co-purchase analysis, integrated into the production environment.

Recommendation engine with audio embeddings for a music-tech platform. We combined CLAP audio embeddings with co-play graph analysis from tracklists into a recommendation engine that leverages both musical similarity and played-together behaviour, "Shazam meets collaborative filtering". Technique: audio embeddings, graph analysis (NetworkX) and vector search (Qdrant).

Document & language AI

AI assistant at a financial services provider in protective guardianship. We built an assistant that automatically classifies inbound email, generates draft replies from client data, and assembles case files. Set up multi-tenant, with client data strictly separated. Impact: substantially less manual processing work. Technique: email classification + retrieval-augmented generation (RAG) on client data.

Analytics, attribution & optimisation

Unified ad analytics and attribution for a marketing agency. We connected lead-funnel data and advertising data from different sources into a single dashboard with cost-per-lead, funnel lead times and ad-to-lead attribution. Technique: cross-source data pipeline + attribution modelling.

Data architecture and stockout detection at an e-commerce operator. We built the data pipeline that brings together webshop, fulfilment and advertising data, with a data-quality audit and automatic stockout detection. Technique: data integration + rule-based and statistical signalling.

Algorithmic trading strategy for a fintech. For a quantitative trader, we built a pairs-trading system with multiple risk buckets, pair selection based on half-life and backtest Sharpe, and a monitoring dashboard. Technique: statistical arbitrage, backtesting and paper-trading validation.

The throughline

What these projects share is not one technique but one principle: the model is half the work, the integration is the other half. The value arises when a prediction, classification or recommendation lands automatically in an existing system and feeds decisions there. An overview of what these applications deliver on average is in our machine learning ROI benchmark.

Recognise one of these challenges in your own organisation? Get in touch, we map your use case, your data and the feasibility.

Frequently asked questions

What kinds of machine learning projects do you do?

From demand forecasting and forecast-bias correction to recommendation and cross-sell engines, document and language AI (NLP/RAG), energy forecasting, anomaly and fraud detection, and algorithmic optimisation. We focus on models that actually reach production and integrate into existing systems, not standalone pilots.

Which sectors do you have experience in?

Among others manufacturing, retail and e-commerce, logistics and supply chain, financial services, energy, automotive and marketing/adtech. The underlying techniques are cross-sector; the value lies in applying them to your specific data and processes.

Do you actually put models into production?

Yes. Most of the value only arises through integration into existing systems, a dashboard, an ERP, a backend. We deliver models that run where the work happens, with attention to operation and monitoring after launch.