Researchers from MIT, Red Hat, and IBM have developed GIFT, an AI framework that substantially improves the automated conversion of two-dimensional designs into precise, executable three-dimensional CAD models. The system addresses a critical bottleneck in engineering workflows: traditional vision-language models often produce simplistic or syntactically flawed CAD code due to limited training data, hindering rapid prototyping and virtual component testing. GIFT, which stands for Geometric Inference Feedback Tuning, utilizes a model-aware data augmentation strategy that allows the AI to learn directly from its own errors. Rather than depending on static, human-curated datasets, the framework evaluates a pre-trained vision-language model through parallel inference runs. It specifically targets mid-success scenarios where the model achieves approximately a fifty percent success rate. When the model generates near-miss code, GIFT automatically corrects the syntax and logic, transforming these failures into functional CAD scripts. These corrected outputs, alongside successful solutions, are compiled into a specialized dataset that retrains the model to handle complex geometric reasoning. The architecture employs inference-time scaling, enabling engineers to direct a specific computational budget toward error correction without the expenses of full model retraining. This dynamic approach yields CAD programs that demonstrate significantly higher geometric accuracy than competing techniques while consuming roughly twenty percent of the standard computational resources. By prioritizing accurate foundational geometry, the generated models reliably pass downstream virtual crash and durability simulations. According to lead researcher Giorgio Giannone of the MIT Design Computation and Digital Engineering Lab, the framework automates what has traditionally required extensive human oversight, turning model failures into continuous training data. The technology streamlines the design iteration cycle, reduces prototyping costs, and helps engineering teams identify optimized configurations that might otherwise remain undetected. Co-senior author Faez Ahmed emphasized that this self-improving capability brings trustworthy AI design tools closer to everyday industrial applications. The research was recently presented at the International Conference on Machine Learning. The development team, which includes contributions from IBM and Red Hat, plans to expand GIFT to optimize manufacturing parameters and scale the system across larger foundation models and broader CAD generation tasks. Funded in part by the MIT-IBM Computing Research Lab, the framework marks a significant step toward autonomous, computation-efficient AI integration in engineering design pipelines.