Mateja Milicevic Is Helping Build the Intelligence Behind Human Behavior Simulations

Mateja Milicevic is a leading research engineer in human behavior simulation. He joined Aaru as the company’s ninth employee during the establishment of the research organization and helped build and scale the organization’s structure. As one of the company’s first three researchers, Milicevic took responsibility for developing and improving the mathematical models that power Aaru’s human behavior simulations. As a founding simulation researcher, Milicevic was critical to the technical foundation of Aaru’s systems, contributing to early research decisions that have influenced everything that followed.

Milicevic focuses on identifying patterns in how human populations respond to products, policies, and changing economic, social, and environmental conditions. He combines traditional statistics with machine learning and artificial intelligence to augment model accuracy and deepen insights. At Aaru, Milicevic owns core components behind the simulations, including: [Insert the names or descriptions of the systems Mateja personally developed or currently owns.]

“I joined when the research team was still very small, so there was nowhere to hide behind an established process,” Milicevic says. “You had to understand the problem, build something that could address it, and take responsibility for whether it worked.”

Milicevic’s central role at an early stage distinguishes his experience from that of a researcher entering a mature technical organization. Milicevic was not assigned to maintain a narrow piece of an already completed system – he built, developed and tested the underlying approach in its nascent stages. His contributions helped improve the capabilities of simulations used to provide strategic insight to large organizations.

The responsibility attached to Milicevic’s work is significant. A human behavior simulation may inform how an organization approaches a market, evaluates a product, or studies the priorities of a population. The research must be handled with discipline because the outputs support decisions with broad ramifications for Aaru’s customers and their target populations. Accuracy cannot be assumed simply because a model produces a clear or confident answer.

“There are many approaches that look promising at first and then fail when you examine them closely,” he says. “A large part of the job is understanding why something failed and finding a stronger way forward.”

Milicevic believes the field is often misunderstood as a simple application of large language models. The common assumption is that researchers ask a model to imitate a person and then treat the answer as a simulation. Aaru’s work involves a more complex research process, drawing from established statistical methods and machine learning techniques alongside newer AI systems to produce insight-rich, population-level patterns at scale.

Statistical and AI methods become valuable tools for studying human behavior when research examines group responses – both to broad conditions and micro-stimuli. Milicevic’s work focuses on larger patterns and population interaction with economic priors. Aaru simulates population behavior without claiming perfect knowledge of every person represented.

“We are looking at what happens across a population,” Milicevic says. “The work becomes useful when the model can show meaningful patterns that help someone understand a larger situation.”

A pilot simulation requested by Ernst & Young (EY) in 2025 provided a landmark proof of concept. EY engaged Aaru to simulate the results of its marquee Global Wealth Survey prior to publishing its 2025 edition. While EY surveys wealth management clients through traditional methodologies to build its report, Aaru’s simulation matched the unreleased data with a 0.90 Spearman correlation – and proved more accurate than EY’s existing data upon outcome-based review.

To Milicevic, examples like the EY benchmark are particularly powerful because they prove the accuracy and value of simulation methodologies against real results and shift discussions away from broad claims about artificial intelligence. A simulation should be judged through the quality of its output, the depth of its insights and the accuracy of its conclusions against observed outcomes. Aaru’s research is constantly improving the methods behind its results.

Milicevic’s mandate has grown and evolved as Aaru has expanded. Since Milicevic joined, the company has more than tripled in headcount, raised a Series A at a reported billion-dollar valuation, and secured strategic industry partnerships. Those developments only increase the importance of Milicevic and his research to the organization and demonstrate how quickly the technical organization around him has grown since he pioneered human simulations at scale as one of Aaru’s earliest researchers.

Growth has not made the underlying problem any easier. Human behavior simulation, as Milicevic describes, demands creative and unusual approaches to model development, and long hours to refine that research. Researchers must be willing to discard ideas that fail even after substantial effort. Progress depends on technical skill, but he places equal importance on persistence.

“The hardest problems do not give you a clean answer just because you worked on them for a long time,” he says. “You have to keep testing, keep adjusting, and be honest about what the system can actually do.”

Milicevic’s interest in predictive models developed through his education in data science and later doctoral study in metagenomics at Michigan State University. He left his PhD program to join Aaru’s research team. The central challenge aligned closely with his long-standing interest in using mathematics to understand outcomes, and an early-stage company with frontier technology appealed to him.

Milicevic now works at the point where abstract research is transformed into a system that organizations can use. His responsibilities include improving model performance, examining failed approaches, and helping shape the research direction behind Aaru’s simulations. Milicevic is a key contributor to Aaru’s population generation models and owns most of the company’s internal benchmarking and evaluation suites. “I do not know which one is harder — developing models to achieve our goals, or evaluating them,” Milicevic says of the day-to-day demands of simulation research.

Milicevic wants to continue pushing the boundaries of simulation research. He is interested in human behavior, but his longer-term vision extends to simulation more broadly. He sees the field as a way to deepen understanding of the world and the patterns that shape it.“

The work is exciting because there is still so much we do not know how to simulate well,” Milicevic says. “Being part of the team that has to figure that out is exactly where I want to be.”