I'm ZHOU Jie, a Microsoft MVP since 2023 and author of the “.NET骚操作” blog. For over a decade I've been bringing OCR, LLMs, media and high-performance computing to .NET — and now even the inference engine itself is pure C#.
// Sdcb.SimdPaddleOCR —— GEMM micro-kernel (illustrative)
static unsafe void Kernel8x8(float* a, float* b, float* c, int k)
{
Vector256<float> c0 = default, c1 = default; // … c7
for (int p = 0; p < k; p++, a += 8, b += 8)
{
var bv = Avx.LoadVector256(b);
c0 = Fma.MultiplyAdd(Vector256.Create(a[0]), bv, c0);
c1 = Fma.MultiplyAdd(Vector256.Create(a[1]), bv, c1);
// … c2 ~ c7
}
Avx.Store(c, c0);
Avx.Store(c + 8, c1); // …
}Now
A ground-up reinvention of PaddleSharp and OpenVINO.NET for the AI era: no Paddle Inference, no ONNX Runtime, no OpenCV — every operator is hand-written C#. 2.0 adds Vulkan / Metal GPU on top of CPU SIMD: 15× end-to-end on an RTX 3080 Ti (medium).
Pure-C# document shadow removal, zero inference-engine deps — compiles to multithreaded WASM, try it in your browser.
Forgot how to split a character? See its roots, strokes and ID code at a glance.
Pure-C# Hunyuan translation on the GPU: 15.8k tok/s prefill on a 3080 Ti, weights on NuGet.
Evolution
Five years, three answers to the same question: how do .NET developers get the best OCR?
Binds the native Paddle Inference runtime
Binds the native Intel OpenVINO™ runtime
The inference engine itself is C#
Open source
.NET bindings for Baidu Paddle Inference — PaddleOCR, PaddleDetection and more, out of the box. One of the most popular OCR stacks in .NET.
A high-quality .NET wrapper for the Intel OpenVINO™ toolkit — fast, frugal inference on CPUs and iGPUs.
Low-level FFmpeg API for .NET generated by CppSharp — decoding, encoding, filters, muxing.
A modern .NET library over GMP and MPFR — arbitrary precision with native speed and .NET ergonomics.
A fast, sandboxed C# runner built for LLM code interpreters over MCP.
A template engine for dynamic SQL, with 100k+ NuGet downloads.
Writing
113 posts on cnblogs and WeChat, from Direct2D in 2013 to hand-written SIMD in 2026. (Written in Chinese.)
Community
Three QQ groups, five WeChat groups and an official account — mostly Chinese-speaking. Questions, ideas, or just lurking — all welcome.
Join the community
From a C++ programmer writing Direct2D in 2013 to a .NET developer hand-writing SIMD inference kernels today. I believe C# can do far more than most people think — every project on this site is an attempt to prove it.
About me