What if C, Python, and Lisp had a child?

Nearly every major data tool in the Python ecosystem is built the same way: a fast C, C++, Rust, or Fortran core wrapped with Python bindings. NumPy, SciPy, Pandas, Polars, PyTorch, xgboost, the pattern repeats. What if you could combine the convenience of Python, with a powerful type system, and the speed of C?

(And maybe without waiting 20 minutes to a couple of hours for Cutlass and Flash Attention to compile)

I am building Tattletale, an AI inference engine, with all the scientific computing primitives needed to reach state-of-the-art performance in terms of plain speed, concurrent serving, multi-GPU support, and embeddability. Single binary, no dependency hell.

The foundation is Nim. Python-like syntax, compiles to C, C++, or JavaScript. Types. Zero-overhead bidirectional FFI with C and C++. A Python bridge that works both ways and can call Python from Nim or Nim from Python. Direct AST manipulation at compile time for metaprogrammingin the Lisp tradition.

This combination means: one codebase generates GPU kernels for CUDA, OpenCL, Vulkan, and WebGPU at compile time via Nim macros.
No multi-stage compilation, ninja or nvcc needed where you deploy your code.
The same source produces a Python-callable library, a C-embeddable binary, and even JavaScript or WASM targets.

This talk covers:

• What Nim looks like in practice
• How the bidirectional Python bridge works — calling Python from Nim and Nim from Python
• Showcasing Nim's Domain Specific Languages: einsum as macros, neural network DSLs
• Compile-time kernel generation: a Nim macro that emits CUDA + Vulkan from the same Nim source
• CuTe Layout Algebra in Nim
• How to seamlessly integrate with LibTorch C+_+ and even catch C++ exceptions.

No Nim knowledge required. Familiarity with FFI or GPU programming helps but is not essential.