Having access to the parton-level kinematics is important for understanding the internal dynamics of particle collisions. Here, we present new results aiming to an efficient reconstruction of parton collisions using machine-learning techniques. By simulating the collider events, we related experimentally-accessible quantities with the momentum fractions of the involved partons. We used photon-hadron production to exploit the cleanliness of the photon signal, including up to NLO QCD-QED corrections. Neural networks led to an outstanding reconstruction efficiency, suggesting a powerful strategy for unveiling the behaviour of the fundamental bricks of matter in high-energy collisions.