Collaboration: Espresso + Orion benchmarks + ANE best practices

[christopherkarani]

Hi mechramc,

Came across Orion while working on Espresso — we're both solving the same problem (direct ANE inference, bypassing CoreML).

Espresso compiles transformers to ANE via MIL and we're seeing 4.76x throughput vs CoreML on transformer inference. Pure Swift, zero dependencies, MIT licensed.

Given the overlap, I think there's value in:

1. Cross-project benchmarks — Espresso vs Orion on the same hardware/models, giving the community a clear comparison
2. ANE best practices doc — we've hit similar MIL compilation gotchas, could co-author a reference
3. Cross-promotion — different audiences might find each project through the other

Would you be open to a benchmark comparison or technical discussion? Happy to run Espresso against your test cases and share results.

Espresso: https://github.com/christopherkarani/Espresso
Paper: [your arxiv link]

Best,
Chris