I shipped an engine for your own UI: three DLLs, CPU, no CUDA required.
I spent a long time building fruit inspection in “Fruit Inspector” — camera, line, control zone, classification. Not everyone wants my exe or my UI. Some people need detect → features → Cat inside their app: C#, Delphi, whatever.
Yesterday I put that out separately: https://github.com/olesha-ai/fruit-inspector-engine
Why
I get this feeling that without CUDA you’re a second-class developer. In this industry you often don’t get a real choice. This package is different: inference on CPU via ONNX Runtime. Simple: drop “onnxruntime.dll” in “lib/cpu”, point init at that path, go. No GPU — bare CPU.
These are not reference libraries and not perfect — I know. But a start is on the table, and that’s what matters to me.
What’s inside (short)
Three DLLs:
“fruit_infer.dll” — NV12 frame → YOLOX boxes (AABB, not Ultralytics OBB).
“fruit_features.dll” — RGB crop → 44 floats (apple pack).
“fruit_cat.dll” — 44 floats → class and score.
Plus headers, models (“det.onnx”, “cat.onnx”, “features.json”), licenses folder — eval plus ORT / YOLOX / LightGBM texts.
Download: Releases → “fruit-inspector-sdk-win64.zip”.
No Nim in your process. Load the DLLs the way you already do.
What’s not there — straight talk
No ready-made C# / C++ sample — only “.h” files.
Get olesha-ai’s stories in your inbox
Join Medium for free to get updates from this writer.
“Fruit_Inspector.exe” with camera and conveyor logic — that’s https://github.com/olesha-ai/fruit-inspector-demo (another repo).
Linux / ARM — Windows x64 only for now.
No RGB crop from NV12 inside the DLL yet — you crop on your side (min 5×5).
No pretending it’s “one button and done.”
License
Non-commercial eval: desk, lab, school, your own experiments — fine. Paid sorting line, OEM, reselling — no. Read LICENSE and ENGINE_EVAL_LICENSE in the repo before “LoadLibrary”.
Who might care
People who write their own camera or bench software and don’t want someone else’s dependency zoo for one inference stack. Anyone willing to read HOST.md once and wire the chain — or just one of the libraries, whatever fits you.
If you’re integrating and writing code against this — you’ll get there. I know you will.
Links
Engine (DLLs, zip): https://github.com/olesha-ai/fruit-inspector-engine
Demo with UI: https://github.com/olesha-ai/fruit-inspector-demo
Feedback — GitHub issues if you actually tried to integrate, not just downloaded.
olesha-ai
