Unlocking value through AI Applied Research Engineering

Andreu Mora · SVP / Global Head of Engineering Data, Adyen

In this day and age, any company — and especially tech companies — needs to find a way to adapt and embrace AI. There’s zero chance that the world and society won’t change because of AI.

On one side, we see how AI labs and the scientific community are spearheading the creation of frontier models and on the other side, how companies clearly see the value of AI but are still pinning down how they can apply these new technologies day to day in a meaningful way, for example in critical / large-scale production flows.

Initiatives are being created and shipped that propel the technology forward in the right direction (e.g. MCP, A2A), but require network effects, outcome guarantees and other challenges to be solved in order to become a full-fledged reality. The world of fintech, and finance in general will change because of AI, and Adyen is exceptionally positioned — given our tech-first mindset, scale and high quality data — to play a primary role in this evolution. We have a great opportunity at being the fintech company that bridges the worlds of AI and finance.

✏️ If you are new to Adyen, Adyen is a global financial technology platform, founded in Amsterdam in 2006, that powers end-to-end payments, data insights, and financial products for many of the world’s leading businesses — including Uber, eBay, and LinkedIn. We are a team of 4500 people who operate in 28 offices worldwide, serve customers in over 115 nationalities, and processed more than USD 1.4 trillion in payment volume in 2024 alone, with a steady growth of ~25% YOY. To this end, we are executing on our AI strategy, which roughly is comprised of three pillars:

1. Adapting our ways of working and productivity tooling to be AI-first. We are doing this using AI products that accelerate our delivery and augment the depth of our work. As examples, we are using conference companions, code companions and AI that helps us search and interact with our knowledge base.
2. Launching AI products: even before the advent of GenAI it was clear that, given enough data presence, AI and Machine Learning are the better approach that guarantees optimal performance and minimal operational cost. We keep moving our product suite to be AI-first. A good example is Adyen Uplift, where an AI trained on a massive dataset decides the course of every transaction by balancing cost, risk and conversion in an optimal outcome and a minimal overhead for the operations team.
3. Engaging with the community and being more exploratory. The AI world is changing rapidly every day and we need to lead the way. We cannot simply remain passive and observe. Exploring, contrary to executing, does not have guaranteed production outcomes and requires ideating, experimenting, iterating and engaging with the AI community in order to learn and reach better outcomes.

In this blog post, I will elaborate on the 3rd pillar, on what Research Engineering means for Adyen, and the initiatives, results, and investments we are making towards that goal.

What do we mean by “research engineering”?

🚀 Focused on impact

The goal of the research engineering team is to bridge the gap between cutting-edge AI research and the use of this research in production applications, by taking research papers and implementing them into robust, scalable, production-ready code. Research engineers do not work alone, they use each other for ideas, sparring, challenging and they also connect with the product engineering teams to bring those ideas to production, after researching and proving their value.

That is fundamentally different from unguided long-term, highly exploratory research that is not anchored on a business goal that would (1) create new products, (2) improve our products or (3) fundamentally change the way finance and companies operate. Instead, we start with a hypothesis that would benefit any of the avenues above and iterate, build and experiment to either accept or reject that hypothesis.

🧪 Access to large scale experiments and data

To that end, Adyen is in the remarkable position of being a medium to small size company with access to a large wealth of data. In order to process all this data, we have invested the last few years in having a top-notch platform to be able to experiment on. This means having the mindset of trying, the engineering to deploy and perform experiments and the scientific mind of rejecting hypotheses. All this over a platform that processed more than 1.4 trillion USD in traffic yearly (2024).

Our data includes not only the transactional information flowing through our API, but also insights from a highly-skilled operational team. This team provides feedback to measure, fine-tune, and reinforce our algorithms.

🏞️ Latched on to the Community

Adyen, like most successful tech companies, is largely built on top of open source. Open source is not only the capacity to use, contribute or share code but also having a mindset of sharing (search “abundance mindset”) where benefitting the world at large, eventually benefits Adyen as well.

Our research engineering team is encouraged to share discoveries, whether it’s through blogs, conferences, or publications. With an insatiable curiosity, the team stays at the forefront of the latest advancements, deeply embedded in the scientific AI research community. They’re also actively connected to the broader ecosystem, including startups, VCs, and AI labs, as well as the product and research engineering teams within our merchant network.

While we collaborate with universities on academic initiatives, the core mission of our research engineering team isn’t primarily to publish academic papers. However, doing so is certainly encouraged if it contributes to the team’s overarching goals.

What are we currently researching?

Here is the list of initiatives we are currently researching:

AI Agentic workflows and Reinforcement Learning for Integrity Risk.

Integrity Risk comprises initiatives such as AML (Anti-Money Laundering), KYC (Know Your Customer), CDD (Customer Due Diligence), CRR (Customer Risk Reviews) which are industry standards designed for financial companies to ensure their activities are affecting the world in a positive way. Given Adyen’s unique positioning of owning banking licenses in the US, EU and UK and acquiring licenses world-wide we are also accountable for the regulators’ input on those areas.

The traditional angle from finance companies would be to employ a sizable team of humans that manually research, analyze and label cases. Before the eruption of LLMs, Adyen already built their own Integrity Risk technology framework, with an AI-first mindset, in which ML models, like Graph Neural Networks, surface cases to human analysts. The framework is designed to optimize precision such that the cases being worked on are worthy of human time, but still requires the deployment of a human team that would eventually spend some time on tasks that have the potential to be automated.

However the advent of LLMs and especially AI agents with access to tools like database queries and web search presents a great opportunity to automate those human-heavy workflows even further. Integrity Risk is an ideal field to introduce research engineering because its stakeholders are internal customers, providing a direct feedback loop from operational teams that are reviewing the cases, which also exposes a great productivity metric which represents how we benefit the company and the world at large.

This data not only provides a way to measure and harness the performance of AI Agents, but it contains a wide set of signals that can be used to accelerate research on Reinforcement Learning (e.g. using PPO orGRPO) to impact the optimization of agent trajectories in highly regulated environments.

We can build on top of the GenAI platform that hosts open source and open weight models which we can contextualize, fine-tune and distill. Last year, we targeted our Customer Support function to mature our tech stack and talent given the nature of the problem (NLP only), but now we are working towards solving more complex tasks. Exploration on top of the public cloud is also possible; in the end what matters is that we evaluate the options available to us as factually as possible to maximize speed and outcome.

It’s a complex, multimodal problem that requires dealing with text, image, search, and complex reasoning , and advancements in this area, with access to a large amount of high-fidelity data, can lead to substantial productivity gains.

AI agents for data analysis.

Agents are a major research avenue for 2025 with significant potential, although a gap still exists for production-ready flows. Especially in the world of finance, many people need to reconcile domain knowledge encoded in manuals, handbooks and other unstructured data sources with other sources of structured data (e.g. transactions, stocks) which typically live in databases or spreadsheets.

Think about doing calculations before the first calculator. Just as the calculator revolutionized our ability to process numbers, trustworthy AI that can reconcile structured and unstructured data, generate insights, leads, recommendations, and spot mistakes would fundamentally transform how we work and spend our time at Adyen, and in the world at large.

We started this research by a very honest initiative: ensuring that we have good evals in place to measure progress but also to help the world solve this problem. To this end, Adyen co-built the DABStep (Data Agent Benchmark for Multi-step Reasoning) benchmark with Hugging Face, which uses over 450 real-world tasks from Adyen workloads to evaluate agent capabilities, specifically requiring reasoning over structured and unstructured data and iterative problem-solving steps. We found that current state-of-the-art reasoning-based agents achieve only around 16% accuracy on DABStep, highlighting the significant need for progress in this area.

Image I: figures for DABStep baselines on current SOTA models.

Foundational model.

This is a highly exploratory project, inspired by the great people of Hyperplane. The hypothesis is that we can leverage the massive scale of our data (both structured and unstructured), Transformer-based architectures and representation learning methods to learn diverse embeddings of entities such as transactions, merchants and “persona” across our platform.

The applications of such a modeling approach are broad for Adyen. We can train different heads simultaneously to solve use cases in the areas of transaction fraud and optimization (connecting that to Uplift), enhancing our graphs, learning representation that help our integrity controls and automate a lot of our workflows while keeping differential privacy.

We have many hypotheses that we are researching and implementing. For example, whether a denoising autoencoders (DAE) would beat the performance of others such as BERT-style architectures.

Image II: Visualization of the representations produced by a DAE, projected via t-SNE. On the left we observe how the input data resembles random noise inside a unit circle. On the right, after feeding this input data to our model we can distinguish structures of the learned representations produced by the model.

Deep learning and reinforcement learning for Uplift.

Adyen Uplift — Adyen’s tool used to balance cost, conversion and fraud for maximum performance and minimum operational load — is already a product heavily driven by AI. In this blog post we dove deep into how it works on the inside and painted the broad strokes of what we are researching.

The research here involves continuous investment, exploring Deep Learning and Reinforcement Learning Off-Policy Evaluation, Policy Gradients, and Causal Inference techniques. The ultimate goal is Multi-Objective Optimization, bringing multiple models together (risk, authentication, authorization, retry, personalization) to achieve better overall performance for the product. The research work here is deployed platform-wide, affecting every single transaction, allowing for massive impact and experimentation potential

🌷 More to come

These are the current initiatives we are working on. However, we have a high-cardinality set of impactful problems to work on that could become promising research engineering avenues.

Building the mindset and the team

Research engineering teams 🤝 Product engineering teams

The research engineering team does not behave like a product engineering team, bound to standard software development practices. As such, they work in a more exploratory, iterative way while staying fixated on business outcomes. This requires curiosity, adaptability and thirst for answers, which, coupled with the current on-going pace of evolution in the world of AI, provides a powerful recipe for progress.

Product engineering teams are also encouraged to innovate within their possibilities and scope. It is not about centralizing all the innovation in one place, innovation should happen everywhere. However, product teams do not always have the bandwidth and some initiatives require joining several teams together. That’s where the research engineering team and product engineering teams partner up to bring these exploratory initiatives into production value.

🏘️ Research hubs

The current research engineering team is based in Amsterdam and Madrid, where the majority of production engineering is based in Amsterdam and Chicago. Given the ecosystem of the Bay Area and the already solidified presence of Adyen in San Francisco, we are also extending the research engineering team to San Francisco, with the goal to connect more deeply with the AI scene and the mindset in Silicon Valley.

🏗️ Building the team and its purpose

The team is also responsible for hiring research engineers, choosing initiatives and also mentoring engineers both in the research and in the production engineering teams in order to experiment and bring value to production.

If you are interested in joining, here are all the resources and open positions.

🛜 Connectedness

Adyen is deeply connected to the top technology companies and the global AI ecosystem through several avenues: (1) by serving tech giants being the financial technology platform of choice for many leading global businesses (e.g. Meta, Uber, Microsoft, Netflix, Google) ; (2) collaborating with AI and tech partners (e.g. Hugging Face); and (3) universities (Adyen funded PhD positions with UvA to research Reinforcement Learning for Uplift).

All in all, our stance on research engineering is not dissimilar to Google’s Hybrid Approach to Research from 2012: strong focus on impact, experimentation at large scale and aim for high outcomes despite high uncertainty. We are, indeed, investing further in research engineering to push the boundaries of our products and financial technology on the world at large.