Since traffic patterns are highly dynamic, ensuring efficient load distribution in network architectures providing multiple paths between a pair of edges is a challenging task. Routing methods like Equal-Cost Multi-Pathing (ECMP) cannot take load state conditions into account during path determination for arising network flows. This behavior can lead to inefficiencies like imbalanced path saturation causing congestion and decreased flow experiences. To consider likely path utilization for advanced routing decisions, machine learning-assisted traffic prediction can be leveraged to estimate and aggregate flow-based load profiles.In this paper, a self-driving system for intelligent flow routing in programmable networks is proposed. A traffic prediction pipeline consuming early and focused metadata views as analysis input provides approximated flow load profiles used to derive probable future path states. Based on these, emerging packet streams are efficiently steered across available path capacities.Depending on the reliability of analysis results in terms of prediction errors, experiments show that, compared to ECMP, using established flow profiles for traffic distribution improves load sharing w.r.t. closer balanced path utilization trends over time.