Veyra: An Exploratory Study of Typed AI Programming Abstractions and Interpreter Overhead

Veyra: An Exploratory Study of Typed AI Programming Abstractions and Interpreter Overhead Veyra is an experimental programming language prototype designed to explore explicit abstractions for artificial intelligence workloads, including tensor operations, partial static shape checking, automatic differentiation, and permission-controlled model loading. This exploratory preprint presents the language design, prototype implementation, functional validation, and descriptive CPU microbenchmarks comparing Veyra with Python/PyTorch. The evaluation includes 31 automated tests, a local Iris classification experiment, and runtime measurements across multiple tensor sizes. A supplementary decomposition experiment investigates the contributions of tensor construction and interpreter execution to observed overhead. The results demonstrate preliminary implementation feasibility while identifying limitations in performance, type guarantees, experimental control, and reproducibility. Research status: Exploratory preprint. Not peer reviewed. Author: Israel Teran Sandoval, Independent Researcher. Manuscript version: 0.6. Evaluated prototype: v0.1.