The project explored how artificial intelligence (AI), combined with synthetic datasets and rule-based models, can support decision-making in early-stage urban development. Using the generative design platform Hektar as a test bed, the team developed, implemented, and evaluated two complementary computational approaches: a deterministic, explainable algorithm and machine-learning models trained on large-scale synthetic data representing over two million urban plot configurations.The deterministic model demonstrated high precision, achieving less than ±5 % deviation between user-defined targets (FAR, SCR) and resulting outputs for 95 % of test plots. The machine-learning work progressed in two stages. In the first stage, a numerical Multilayer Perceptron (MLP) outperformed a convolutional neural network (CNN) in predicting Site Coverage Ratio (SCR) from compact geometric descriptors. In the second stage, a refined model predicted probability distributions of SCR outcomes, reflecting the stochastic generation of building configurations in the updated Hektar system. Together, these methods established a reproducible workflow that translates user goals into valid spatial outcomes while introducing probabilistic reasoning to early-stage planning.The project demonstrates that small, task-specific AI models can be computationally efficient while delivering substantial benefits. By improving the precision of density and form assessments, such models can contribute to reduced material use, more efficient land allocation, and lower climate impact in the built environment.All predictive models, datasets, and documentation are published openly on GitHub to support further research and industry adoption. By shifting from form generates data to data generates form, the project outlines a scalable pathway toward prescriptive, data-driven urban planning tools capable of supporting more sustainable, evidence-based decisions