Miroljub Mihailovic

My focus lies in two critical areas: executing unit tests and performing code refactoring for ROS-Neuro. These concerted efforts are aimed at bolstering reliability, validating functionality, and optimizing the codebase for enhanced maintenance and scalability.

I spearheaded the development of Prioritas and played a pivotal role in overseeing and contributing to the creation of Kassandra. Prioritas is a cutting-edge system that speeds up the generation of Suspicious Activity Reports (SOS) by ranking AML alerts based on their potential significance. This innovation reduces SOS processing time by 60% to 80% and minimizes errors in predicting low-risk alerts. Kassandra, a powerful AML module fully based on artificial intelligence, uses pattern recognition to uncover hidden anomalies in transaction data. It boasts high precision, with 90% of subjects flagged for further investigation and 52% resulting in SOS reports. Together, these solutions have greatly enhanced our AML and transaction monitoring capabilities, ensuring efficient and accurate operations in the evolving landscape of financial security and compliance.

• Autonomous Docking Manoeuvre Testing – 2025
• 4th Symposium on Space Educational Activities – 2022 (Abstract)
• 72nd International Astronautical Congress – 2021 (Paper)

Within the project, I played a crucial role in designing and implementing the Onboard Computer System (OBCS) of the CubeSat. The primary Raspberry Pi processes sensor data (tri-axial accelerometer, gyroscope, tri-axial magnetometer, proximity sensor) and computes trajectories for maneuvers, sending commands to the actuators. The secondary unit, based on Arduino, provides real-time control of the actuators. We utilized the Robot Framework and the Robot Operating System (ROS) to manage data and maneuvers. My contribution ensured reliable control of the onboard sensors and actuators, contributing to the success of the satellite missions.

Developed an RL policy integrated into a shared teleoperation system. The agent navigates unknown environments with obstacle avoidance, tested in simulations and teleoperated real-world scenarios. Compared RL to existing rule-based policies for efficiency and safety.

Built a neural network for boat recognition with custom dataset and selective search algorithms. Preprocessing improved region proposals. Non-maximum suppression ensured efficient grouping. Successfully identified and bounded boats in varied image conditions.

Implemented multiple approaches: - Exact models (MTZ, Benders, CPLEX) - Heuristics (Matheuristics, Constructive, VNS, Tabu, Genetic Algorithms). Compared performance and applicability for different TSP instances.