How do you apply state-of-the-art machine learning to mission-critical processes when you’re limited by your enterprise architecture or strict rules around security and stability? In this article, Emilio Oldenziel, Lead AI at Eraneos, discusses what he’s learned working for one of Europe’s largest railway operators and how this can be applied in other industries such as banking, government and healthcare. He explains how the team managed to scale up an academically proven solution and how they balanced state-of-the-art technology with the need for human oversight.
Identifying the challenge: A need for innovation in rail operations
Rail operators worldwide face the challenge of maintaining punctuality and operational efficiency while managing vast and complex networks. Balancing the demands of passenger services and cargo transportation with the availability of staff and equipment requires significant resources and precise coordination. State-of-the-art AI, and specifically Machine Learning models, offer a transformative solution to these challenges. By leveraging these advanced technologies, rail operators can enhance their decision-making processes, optimize scheduling, and predict and mitigate potential disruptions before they occur. This proactive approach not only improves service reliability but also helps rail companies stay ahead in a competitive market. As operational efficiency becomes a critical differentiator, the adoption of AI and Machine Learning is increasingly seen as essential for maintaining a competitive edge in the rail industry.
“We saw that the organization had been running various proof-of-concept initiatives with thesis students, which clearly demonstrated the potential of these technologies. However, while these projects showed promise, they lacked the necessary scale and technology to be implemented in a practical, industrial-scale setting. The challenge was to take these insightful, academic-level experiments and develop them into robust, scalable solutions that could be fully integrated into the daily operations of the rail network.
At the same time, the company was hesitant to adopt new technologies that would interact with a mission-critical process such as planning due to concerns about security and stability. The risk of system outages was, for example, a critical concern, as even a minor disruption in the planning systems could lead to significant delays across the entire network. We had to prove that our solution would adhere to the strict rules and standards of their enterprise architecture,” explains Emilio Oldenziel, Lead AI at Eraneos.
Adopting ML Ops to ensure solution quality
The adoption of ML Ops (Machine Learning Operations) was a crucial aspect of the project’s success and vital in ensuring the quality of the machine learning solution. The tools and processes involved in ML Ops facilitated the integration of machine learning models into the existing operational framework of the rail network. This was a critical step in moving from the academic and experimental phase to real-world application. Ultimately, implementing ML Ops also allowed for the creation of a robust infrastructure that could scale with the project’s needs.
“ML Ops was not just a technical choice but a strategic approach to ensure that the ML solutions could be trusted, managed, and scaled within the mission-critical operational context of the rail network operator.”

This approach facilitated the transition from experimental models to practical, scalable solutions that could be integrated into daily operations while adhering to strict safety and quality standards.
Following the completion of this project, the lasting impact was recently reaffirmed when one of the team members published a research paper. The paper builds on the approach and solutions developed during this project, enabling further optimization of the rail system.”
To handle the vast datasets and complex predictions required, the team utilized advanced algorithms such as reinforcement learning, which is well-suited for decision-making processes where outcomes are uncertain. In rail operations, factors such as sudden weather changes, unexpected equipment failures, or unplanned passenger surges can unpredictably impact schedules and require dynamic, real-time responses. The system could learn from previous outcomes and adjust strategies in real-time, providing optimal responses to varying conditions. This adaptive approach was crucial for managing the dynamic nature of rail networks, where delays can have cascading effects.
Winning the Data Analytics Race in Transport & Logistics
Learn how Data Analytics is empowering companies in the Transport & Logistics sector to remain relevant and gain competitive advantage.

Overcoming challenges: Balancing technology with human oversight
Implementing these advanced solutions required navigating several significant challenges. One of the primary hurdles was the restriction on using certain cloud-based capabilities due to security and operational concerns. This led the team to opt for open-source tools, such as MLflow, to manage ML models.
“Due to the restrictions that come with the enterprise architecture of these mission-critical operations, using various cloud solutions wasn’t an option. Security and operational stability were paramount, so we opted for MLflow, an open-source alternative. This allowed us to maintain control over the technology and run the solution on-premises, tailored to the specific needs of the company. It forced us to be creative and discover how we could achieve the desired results in a different way.”
Another critical aspect was that despite the use of advanced AI, human involvement remains crucial, particularly in scenarios where decision-making directly impacts safety and operational reliability. This is described as a “human-in-the-loop” approach, where humans oversee and validate the decisions made by AI systems. This ensures that AI recommendations align with safety standards and operational protocols. Involving humans in the decision-making loop was not only a safeguard for reliability but also a means to build trust and acceptance among end-users and company leadership by transparently demonstrating the system’s reliability and effectiveness.
“One of the strengths of Eraneos is that we’re not only able to develop an approach where humans can work with state-of-the-art technology but also be able to automate routine tasks, freeing human operators to focus on more complex and critical decision-making. This approach allows human resources to be utilised for higher-value tasks that require ethical judgment and nuanced understanding, which are beyond the current capabilities of AI. By showcasing the transparency and effectiveness of the system, we also helped build confidence among the people involved and ensure that there is no reputational risk associated with automated decision-making.”
Future applications: Expanding AI to other mission-critical sectors
“We see tremendous potential for applying our approach to other industries that have mission-critical processes, such as banking, healthcare, and government operations. These sectors face strict security and operational requirements, similar to what we dealt with in the rail industry. By leveraging state-of-the-art machine learning solutions, we can help these industries optimize their operations, improve decision-making, and stay ahead of the competition.
However, it’s not just about implementing technology—it’s about ensuring these solutions are secure, scalable, and can be trusted by human operators. This will require a careful balance of innovation, compliance, and user acceptance, which is both a challenge and an opportunity.”
For Eraneos, the challenge of implementing these types of solutions across diverse industries is an exciting prospect, promising both technological advancement and real-world impact. The goal is clear: to empower industries to thrive in an increasingly competitive and data-driven world.
AI: Are you ready to go beyond Proof of Concept?
Want to know all about overcoming the challenges of implementing AI and achieving cross-functional success? Our AI and Innovation newsletter has you covered.
