Since 2023, Faktion and TOMRA Food have built the strategic backbone of TOMRA Food's deep learning at a global scale: one shared platform, connected to the sorting machine fleet, where teams manage data, train models and release them to machines worldwide.
Most deep learning in food sorting stops at a lab experiment or a single application. TOMRA Food applies it to the full range of food sorting, from fresh produce to processed foods, at production scale.
Connectivity, data and deep learning make that possible. The platform connects model development with controlled releases to machines in the field. Each customer gets its own models and decides when to install an update and whether to share its data to improve them.
Faktion built it as an "AI Product Studio, as a Service" engagement: commissioned by TOMRA Food, for TOMRA Food, and used by TOMRA Food alone.
TOMRA Food, a subsidiary of TOMRA Systems ASA, is the global leader in sensor-based sorting machines for the food industry. The company wanted to take deep learning out of R&D work and turn it into a governed capability that internal teams could use. To develop it, TOMRA Food brought in Faktion, an AI consultancy with more than a decade of experience turning complex AI into working products.
Together, the two teams designed and built a global data centralisation and MLOps collaboration platform, connected to TOMRA Food's machine fleet. It lets data scientists, application specialists and sales engineers work through the full model lifecycle: data collection and annotation, training, evaluation, release and deployment tracking.
The opportunity: deep learning at the scale of a global fleet
Applications sit at the heart of TOMRA Food's business. Each application covers one product type, from fresh produce to processed foods, and each needs its own classifiers tuned to how that product behaves and how it is sorted.
The scale behind that is considerable: TOMRA Food serves customers in many countries, with a wide range of applications and product categories. Its machine families combine cameras and sensors for RGB, infrared, X-ray and spectral imaging.
As the market moved from traditional machine learning and pixel-based segmentation towards deep learning, an opportunity opened up. A rules-based sorter is only as good as the defects someone thought to describe in advance. Deep learning keeps going: it improves for as long as it gets new examples to learn from.
For TOMRA Food, with a connected global machine fleet spanning multiple regions, products and applications, deep learning became a point of difference. Going from one machine to the whole fleet meant scaling:
- Global cooperation
- Structured datasets
- Consistent model management
- Annotation that keeps up with the volume
- Controlled deployment.
So the platform had to work as a product for the whole organisation, with research, application and service teams all inside it.
A platform for the whole organisation
TOMRA Food and Faktion designed and built that platform together. It supports the full lifecycle of a classifier for each machine, application and region, from data ingestion and annotation to training, evaluation, controlled release and deployment tracking.
In production, teams manage training datasets, millions of annotations made by people and automated ML pipelines, thousands of training runs and hundreds of released models. Each model reaches the field through a controlled release process, and customers decide for themselves when to install an update.
The people using the platform matter as much as what it does. The AI behind it is complex, but the interface is simple enough to use without a data science background. So application specialists and sales engineers take part alongside the data scientists. Customers also choose whether to share their data to improve the models.
We set out to build more than better deep learning models. We wanted a platform where teams around the world improve those models themselves. By managing data, datasets and models in one central environment, we give data scientists, application engineers and service engineers access to the same powerful toolset. That speeds up the development of AI for industrial food-sorting machines and makes it easier to share knowledge and improvements across teams worldwide.
Seppe Van Isterdael, Project Manager Faktion.
Built in partnership, adopted globally
TOMRA Food teams in Dublin, Leuven and New Zealand were involved in building the platform from the start. Through frequent demos and working sessions, data scientists, application specialists and sales engineers helped design it around their daily workflows.
Adoption followed, those same teams now use the platform every day, in different time zones. Releasing a model takes days instead of months, so TOMRA Food can cover more applications and product categories without adding manual work.
Faktion worked as a true extension of our team. Their experienced data scientists and full-stack developers helped bridge two worlds that are often difficult to connect, while keeping communication close and delivery on track.
Alexander Damen, Team Lead AI at TOMRA
What makes TOMRA so significant is its strategic importance on a global level. This is the core of TOMRA's business: sorting food to the next level through connected machines running on gigantic amounts of global data.
And what we have built here is not generative AI or agentic AI. It is the pure power of classic deep learning, engineered to be state of the art and genuinely robust. Over a years-long partnership, TOMRA has had the vision and the courage to keep pushing that forward, at the right time, in the right place.
Our cooperation with TOMRA FOOD is Faktion at its best: AI Product Studio as a Service. Together, we built a globally connected Vision AI ecosystem, combining custom AI, software engineering and MLOps into a single operating platform. A true dream cooperation.
Bart Baeyens, CEO Faktion
.avif)
