Lucas Reymond

Panipro SA is a Swiss software company that develops digital solutions for bakeries of all sizes, from small independent businesses to industrial bakeries. Its products help clients manage production planning, inventory, sales analysis and operational processes by integrating data from point-of-sale systems, production facilities and business management software. During my time at Panipro SA, I was responsible for developing an end-to-end machine learning solution for demand forecasting in bakeries. The objective was to predict product sales several days in advance so that clients could improve production planning, reduce unsold goods and limit food waste.

The project used up to twelve years of historical daily sales data covering a large number of products and stores. I enriched the data with weather observations and forecasts from MeteoSwiss, public holidays, school vacations, closing days and client-specific events. My colleagues helped me understand the existing FoodSystem environment, identify the relevant data sources and clarify the operational needs of bakery managers and production teams. However, I developed the statistical, machine learning and software components of the project independently, as I was the only person in the team with a background in statistics and machine learning.

Machine Learning and Artificial Intelligence

I designed and implemented the complete modeling workflow, from data preparation and feature engineering to model evaluation and prediction generation. I tested several forecasting approaches, including NeuralProphet, LightGBM, XGBoost and TabFM, in order to compare different model families and determine which methods were best suited to the characteristics of the available data.

I created features representing trends, seasonality, recent sales behaviour, calendar effects, weather conditions, holidays and business events. I also investigated how sales patterns differed between products, stores and periods of the year. Because some products had long and stable histories while others had sparse, irregular or highly seasonal demand, the project required more than applying a single model with a fixed configuration.

I studied NeuralProphet in depth, including its trend, seasonality, autoregressive and future regressor components. I compared the behaviour of global and product-specific models, experimented with different lag structures and forecasting horizons, and implemented procedures for model training, saving, loading and incremental retraining. I also introduced automated hyperparameter optimization using Optuna and developed validation workflows to compare model configurations before deployment.

A significant part of my work consisted of diagnosing model behaviour rather than simply producing predictions. I analysed why some configurations failed, investigated inconsistencies between library documentation and implementation, examined the influence of individual regressors and tested whether additional information, such as weather variables, genuinely improved predictive performance.

Statistics and Model Evaluation

I defined the statistical methodology used to assess the quality and reliability of the forecasts. I implemented several performance metrics, including MAE, RMSE, SMAPE, WAPE, mean bias and relative bias, because no single metric was sufficient to describe performance across products with very different sales volumes.

I designed chronological validation and backtesting procedures to ensure that the models were evaluated on future periods rather than on randomly selected observations. I studied the impact of training-window length, model retraining frequency, recent-observation weighting and different feature sets on out-of-sample performance.

I also worked on prediction uncertainty and explored methods such as quantile regression and conformal prediction. More generally, I tried to keep a scientific approach throughout the project: formulating hypotheses, designing controlled comparisons, checking data quality, analysing errors and avoiding conclusions based only on visually attractive forecasts.

This statistical work was important because the final objective was not simply to obtain a model with a low average error. The forecasts also had to remain stable, interpretable and useful for products with different scales, sales frequencies and operational constraints.

Beyond forecasting observed sales, I also developed a methodology to estimate true customer demand. In retail, recorded sales are not always equal to demand because products may be unavailable or sold out before all potential demand has been met. I designed statistical procedures to estimate this hidden demand from historical sales, production quantities and operational information, allowing the forecasting models to learn from a more realistic representation of customer demand rather than from censored sales alone.

Software Development and Production Deployment

In addition to the modelling work, I developed the project as a complete software application rather than as a collection of experimental notebooks. I progressively transformed the initial prototypes into a modular Python codebase supporting data preparation, training, retraining, hyperparameter tuning, backtesting and prediction.

I developed pipelines to retrieve and preprocess sales data, business events, closing days, historical weather observations and future weather forecasts. I worked with SQL Server data and integrated the forecasting outputs with the file formats and processes used by Panipro’s existing software.

I implemented command-line workflows allowing the same codebase to run in different modes, including training, prediction, model evaluation and hyperparameter optimization. I also added model persistence, execution logs, reproducibility controls, CPU resource management and automated generation of diagnostic plots and result files.

To make the results accessible to users without Python knowledge, I developed an authenticated Streamlit dashboard. The application allowed users to select stores and products, visualize historical sales and future predictions, compare forecasts with actual sales and review performance metrics. I later extended the dashboard to support several languages and different types of business data, including product quantities, revenue and transaction counts.

I also contributed to the operational deployment of the system. I packaged the Python applications as Windows executables, configured automated prediction tasks, deployed the forecasting engine and dashboard on test and client servers, and created detailed technical documentation for installation, configuration, maintenance and troubleshooting.

Collaboration, Communication and Client-Facing Work

Although I developed the forecasting solution independently, the project required close collaboration with software developers, my manager and clients. The developers had strong knowledge of the company’s products, databases and technical environment, while I brought the statistical and machine learning expertise. A major part of my role was therefore to create a bridge between these areas.

I regularly explained modelling choices, performance metrics and technical limitations to colleagues who did not have a background in statistics or machine learning. I also adapted the architecture of the project to fit the constraints of the existing software and worked with developers to define how data and predictions should be exchanged between the forecasting system and FoodSystem.

I participated in client meetings with bakery managers, production managers and company directors. During these meetings, I presented the progress of the project, demonstrated the forecasting interface, discussed the results and gathered feedback about how predictions would be used in practice. These exchanges helped me better understand the operational consequences of forecasting errors and ensured that the solution remained relevant to real production workflows.

Initiative and Leadership

Because I was responsible for the project from its early experimentation phase to deployment, I also took initiative in defining its technical direction. I identified several opportunities to improve the original scope, discussed them with my manager and implemented the ideas that were considered valuable.

These improvements included expanding the evaluation framework, adding additional model families, introducing automated hyperparameter optimization, developing more robust validation procedures, improving the treatment of weather and event data, adding prediction uncertainty, automating model execution and creating a user-facing dashboard.

I also proposed changes to make the project easier to deploy and maintain, such as restructuring the code, separating configuration from implementation, packaging the applications and documenting the full production workflow. This required balancing model quality, computing time, maintainability, client requirements and the constraints of the company’s existing infrastructure.

This experience gave me the opportunity to take ownership of a complete AI project and to make decisions across machine learning, statistics, software engineering and product development. It also taught me how to lead a technical project without formal authority: by identifying problems, proposing practical improvements, discussing trade-offs with managers and developers, and turning agreed ideas into working solutions.

Rather than only implementing assigned tasks, I actively shaped the technical direction of the project. Many of the project’s capabilities were ideas that I proposed myself after identifying limitations in the existing solution. Once discussed and validated with my manager, I designed and implemented these improvements independently.

Neuchâtel, since June 2025