Developing effective machine learning pipelines requires appropriate algorithm selection and hyperparameter tuning tailored to dataset characteristics. Traditional AutoML systems rely on exhaustive search or computationally expensive optimization, which is infeasible in many real-world settings. This paper introduces MetaFlow, a metadata-driven AutoML framework that selects models using dataset properties combined with heuristic and constraint rules. Given metadata sample size, feature count, and class imbalance, MetaFlow eliminates unpromising algorithms before training and concentrates computation on the most suitable candidates. Experiments on four real-world datasets show that metadata-conditioned constraints prune the candidate search space by 67% without sacrificing predictive competitiveness.