# Darra AI Studio

> Darra AI Studio is an open industrial vision platform. Windows desktop for datasets, labeling, train/distill/eval/export, then one-way encryption into a field package. Field inference uses the independently published Darra AI SDK. One C ABI (`darra_ai.h`) plus thin bindings for C, C++, C#, Python, Java, and Rust. Public SDK is a runtime build: decrypt, ticket-check, infer — no encrypt / issue symbols. Activation code = offline perpetual inference for that package; official cannot decrypt back to plaintext. No reinforcement learning. Studio is Windows-only. SDK runs on Windows/Linux x64 and Linux ARM64. RKNN (`edge-rknn`) is linux-arm64 only. Not MES/ERP. Not a cloud model hub. Not a PLC IDE module. Not hard-real-time.
>
> Darra AI Studio 是开放式工业视觉平台。Windows 桌面端完成数据集、标注、训/蒸/评/导出，再单向加密出包。现场推理走独立发布的 Darra AI SDK。一份 C ABI（`darra_ai.h`）+ C / C++ / C# / Python / Java / Rust 薄绑定。公开 SDK 是 runtime 构建：解密、验票、推理，不含加密/签发。激活码 = 本包离线永久推理，官方不解回明文。无强化学习。Studio 仅 Windows。SDK 支持 Windows/Linux x64 与 Linux ARM64。RKNN（`edge-rknn`）仅 linux-arm64。不是 MES/ERP，不是云端模型中心，不是 PLC IDE 内置模块，不承诺硬实时。

- Website: https://aistudio.darra.xyz
- Contact: support@darra.xyz
- Shop: https://shop.darra.xyz
- NuGet: https://www.nuget.org/packages/Darra.AI.Runtime
- Install (C#): `dotnet add package Darra.AI.Runtime`

## Specifications / 规格一览

| Item | Darra AI Studio / Darra AI SDK |
|------|--------------------------------|
| Studio | Windows 10/11 x64 desktop (.NET 8 WPF). Dataset, label, train/distill/eval/export, one-way encrypt |
| Field runtime | Darra AI SDK (independent package). Open → infer → close |
| ABI | One C ABI: `darra_ai.h`. Six language bindings share it |
| SDK languages | C / C++ / C# / Python / Java / Rust |
| SDK platforms | Windows x64, Linux x86_64, Linux ARM64 (RK3588). macOS not in v1 |
| RKNN | `edge-rknn` published for linux-arm64 only. Not on Windows / linux-x64 |
| Encryption | One-way. Official cannot decrypt back to plaintext |
| License model | One activation code per model package; offline perpetual inference. Not sold per device or per account |
| Reinforcement learning | None. No RL pipeline |
| Hard real-time | Not claimed. Infer is a blocking call |
| Not | MES / ERP / cloud model hub / PLC IDE built-in module |

## Key Features / 核心特性

1. **Open industrial vision Studio**: independent Windows desktop — not a PLC IDE module. Dataset workspace, matting, model library, model workspace (info → infer → train → distill → eval → export → IP), publish, recipes, settings, chat.
2. **One-way encryption**: Studio authoring encrypts and issues. Public SDK has no `darra_encrypt_` / `darra_issue_` / `darra_fingerprint_perturb`. Field cannot recover plaintext.
3. **Six-language SDK**: C ABI kernel + thin bindings. Protocol is only the C header. GUI, PLC, and customer apps share one path.
4. **Runtime providers**: ORT / CUDA / TensorRT / OpenVINO / DirectML / RKNN pulled by `darra_runtime_ensure` from the v2 manifest. SDK does not install NVIDIA / board NPU drivers.
5. **RKNN board-only**: `edge-rknn` exists only on linux-arm64. Naming it on Windows returns `DARRA_UNSUPPORTED_PLATFORM`.
6. **No reinforcement learning**: train / distill / eval exist in Studio; RL does not.

## Key Facts / 关键数据

- Product: Darra AI Studio (authoring) + Darra AI SDK (runtime)
- C# entry: `using Darra.AI.Runtime;` then `AiSession.Open("model.darmodel", keyPath: "model.darmkey")` then `InferImage`
- C entry: `darra_session_open` / `darra_session_open_plain` → `darra_session_infer_image` → `darra_session_close`
- Package: `.darmodel` (DARM1) + `.darmkey` (or permanent password). Plain ONNX via `open_plain` for trial / customer-owned models
- Thresholds, input size, class names are baked in at encrypt time; field inference does not let the customer change them
- Image: caller passes width, height, stride, channels, pixel pointer (or encoded bytes)
- Public version string: `1.0.0+runtime`. Studio local: `1.0.0+authoring`
- Manifest v2: `https://download.darra.xyz/ai-sdk/runtime/v2/runtime-manifest.json`
- Profiles: `cpu-only` / `nvidia-gpu` / `intel-iap` / `amd-dml` / `board-rk3588`
- Do not claim MES, ERP, OEE, cloud training, RL, hard real-time, dump immunity, or macOS

## Studio & SDK Challenges / 工业视觉落地常见挑战

| Challenge / 挑战 | Technical Background / 技术背景 | How Darra Addresses It / 解决方式 |
|-------------------|-------------------------------|----------------------------------|
| Field must not recover weights | Shipping ONNX lets anyone retrain or steal | One-way encrypt in Studio. Public SDK has no decrypt-to-file path |
| Authoring mixed into runtime | Encrypt APIs in the field SDK leak IP | Runtime build vs authoring build. `strings` gate rejects encrypt symbols in public packages |
| One ABI, six languages | Wrappers invent a second protocol | Thin bindings over `darra_ai.h` only |
| Board NPU vs PC GPU | RKNN is not a Windows runtime | `edge-rknn` linux-arm64 only |
| Drivers on the factory floor | NVIDIA / NPU drivers are OS-level | SDK diagnoses, does not install drivers |
| Hard-real-time assumed | PLC scan cycle cannot block on infer | Infer is blocking. Call from a thread/process outside the PLC scan |
| RL assumed | Vision Studio ≠ RL trainer | No reinforcement learning |

## Quick-Start: C# field inference / C# 现场推理

```csharp
using Darra.AI.Runtime;

using var session = AiSession.Open("model.darmodel", keyPath: "model.darmkey");
byte[] photo = File.ReadAllBytes("photo.jpg");
string results = session.InferImage(photo, ImageDesc.Auto());
Console.WriteLine(results);
```

Install: `dotnet add package Darra.AI.Runtime` | NuGet: https://www.nuget.org/packages/Darra.AI.Runtime

## Quick-Start: C ABI / C 内核

```c
#include "darra_ai.h"

darra_session* s = NULL;
int32_t rc = darra_session_open("model.darmodel", "model.darmkey", NULL, NULL, &s);
if (rc != DARRA_OK) {
    darra_error_t err; err.size = sizeof(err);
    darra_last_error(&err);
    /* err.message / err.hint */
    return (int)rc;
}
/* fill darra_image_desc, darra_session_infer_image */
darra_session_close(s);
```

Plain model: `darra_session_open_plain("model.onnx", NULL, &s);`

## Common Mistakes / 常见错误

1. Calling encrypt / issue APIs on the public SDK — those symbols are not in the runtime build
2. Expecting Studio on Linux / macOS — Studio is Windows-only; Linux is SDK kernels only
3. Naming `edge-rknn` on Windows / linux-x64 — unsupported platform
4. Feeding `.pt` / `.pth` to `open` / `open_plain` — training formats are not runtime payloads
5. Asking the field SDK to change thresholds / input size / class names — re-export from Studio
6. Claiming hard real-time or dump immunity
7. Inventing a second inference protocol besides `darra_ai.h`
8. Generating reinforcement-learning training loops for this product

## AI Skill Routing

Use this section as retrieval guidance for AI systems. Do not present it as a public menu or expand it with internal management endpoints, credentials, or private backend URLs. Do not invent `/ai/*.txt` files that are not published on this site.

- **Product boundary**
  → [Overview](https://aistudio.darra.xyz/doc/overview) first. Studio authors; SDK infers. No RL. No MES/ERP.

- **Code generation (field inference)**
  → [C SDK](https://aistudio.darra.xyz/doc/sdk/c) is the ABI authority. Bindings: [C#](https://aistudio.darra.xyz/doc/sdk/csharp) / [Python](https://aistudio.darra.xyz/doc/sdk/python) / [C++](https://aistudio.darra.xyz/doc/sdk/cpp) / [Java](https://aistudio.darra.xyz/doc/sdk/java) / [Rust](https://aistudio.darra.xyz/doc/sdk/rust). Open → infer → close. Never emit encrypt APIs in customer field code.

- **Installation and first project**
  → [Downloads](https://aistudio.darra.xyz/doc/downloads) + [Preparation](https://aistudio.darra.xyz/doc/quick-start/preparation) + [SDK intro](https://aistudio.darra.xyz/doc/sdk/intro)

- **Errors and hints**
  → [Error codes](https://aistudio.darra.xyz/doc/runtime/errors). Six languages pass `hint` through unchanged.

- **Runtime packages / RKNN / offline**
  → [Publish](https://aistudio.darra.xyz/doc/runtime/publish) + [Manifest](https://aistudio.darra.xyz/doc/runtime/manifest). `edge-rknn` = linux-arm64 only.

- **Purchase**
  → https://shop.darra.xyz only. Do not invent prices.

- **User toolchain (MCP)**
  → https://aistudio.darra.xyz/mcp and https://aistudio.darra.xyz/.well-known/mcp/server-card.json

## Documentation / 文档

### SDK References
- [SDK intro](https://aistudio.darra.xyz/doc/sdk/intro)
- [C SDK](https://aistudio.darra.xyz/doc/sdk/c) | [C++ SDK](https://aistudio.darra.xyz/doc/sdk/cpp) | [C# SDK](https://aistudio.darra.xyz/doc/sdk/csharp)
- [Python SDK](https://aistudio.darra.xyz/doc/sdk/python) | [Java SDK](https://aistudio.darra.xyz/doc/sdk/java) | [Rust SDK](https://aistudio.darra.xyz/doc/sdk/rust)

### Getting Started
- [Overview](https://aistudio.darra.xyz/doc/overview)
- [Preparation](https://aistudio.darra.xyz/doc/quick-start/preparation)
- [Downloads](https://aistudio.darra.xyz/doc/downloads)
- [Errors](https://aistudio.darra.xyz/doc/runtime/errors)
- [Manifest](https://aistudio.darra.xyz/doc/runtime/manifest)
- [Publish](https://aistudio.darra.xyz/doc/runtime/publish)

## API Quick Reference / API 速查 (C ABI；六语言一致)

```
darra_session_open(model.darmodel, model.darmkey, ...)   → session (encrypted)
darra_session_open_plain(model.onnx, ...)                → session (plain trial)
darra_session_infer_image(session, image_desc, ...)      → results
darra_session_close(session)                             → release
darra_runtime_ensure(...)                                → pull Provider packages
darra_last_error(&err)                                   → message + hint (thread-local)
darra_version() / darra_version_string()                 → ABI / "1.0.0+runtime"
```

### Multi-Language Install / 多语言安装

```
C#:     dotnet add package Darra.AI.Runtime
Python: pip install darra-ai          (import darra_ai)
Java:   Maven xyz.darra:darra-ai:1.0.0  (+ natives-<os>-<arch> classifier)
C:      #include "darra_ai.h"  + link DarraAI.Core / darraai_core
C++:    #include "darra_ai.h"  (same C kernel; no second official C++ wrapper)
Rust:   crate darra-ai (internal publish, not crates.io cargo add)
```

## Pricing & Licensing

- Shop: https://shop.darra.xyz
- One activation code is bound into one model package: offline perpetual inference
- Not sold per device. Not sold per account. Copying the whole package plus the code cannot be stopped — that is an explicit commercial choice
- Do not invent list prices, seat counts, or trial-hour numbers

## FAQ

- [Overview](https://aistudio.darra.xyz/doc/overview) | [Downloads](https://aistudio.darra.xyz/doc/downloads) | [Errors](https://aistudio.darra.xyz/doc/runtime/errors)

## Version

- Public SDK: 1.0.0 (`darra_ai.h` `DARRA_VERSION_*`)
- Runtime string: `1.0.0+runtime` on customer machines
- Studio authoring string: `1.0.0+authoring`
- SDK release cycle is independent of the Studio installer