EVREN SDK released: an official Python client for computer vision

EVREN SDK released: an official Python client for computer vision models

EVREN SDK, the client library for the computer vision side of EVREN, a platform developed under Türkiye’s Presidency of Defence Industries (SSB), is now available on PyPI as evren-sdk. The first release, 0.3.0, was uploaded on 14 March 2026, and the package matured on 12 April 2026 (UTC) with versions 0.5.3, 0.6.0, 0.6.1 and 0.7.0 published back to back. It ships under the Apache-2.0 licence and requires Python 3.10 or later.

This news piece is written for software teams that want to catch defects from a production-line camera, count objects with a field device or add image classification to an existing business application. We first summarise what the package documentation states explicitly, then share our own technical commentary on enterprise integration. Aksiyon Soft has no partnership with SSB or EVREN; this is independent industry news.

In short

• What: the official client for running inference from Python against vision models trained on EVREN; package name evren-sdk.
• Releases: seven versions on PyPI between 0.3.0 (14 March 2026) and 0.7.0 (12 April 2026).
• Tasks: object detection, classification, segmentation, oriented bounding boxes (OBB) and keypoints.
• Edge mode: real-time inference on GPU-free Raspberry Pi boards, laptops or industrial PCs, with the computation running on EVREN’s GPU cluster.
• Infrastructure: according to the documentation, a single predict() call is backed by 8× NVIDIA A6000 GPUs.

What is EVREN SDK and what does it do?

The package description presents the SDK as the official tool for running inference from Python against computer vision models trained on the EVREN platform. In other words, training happens on the platform itself; the SDK is how you call a trained model from your own code. Authentication uses a key generated under Settings → API Keys on the platform, prefixed with evren_.

Basic usage takes three lines: create a client, pass a model name and an image, read the returned predictions. A model can be referenced as user/model-name (latest version), with a version tag such as :v2.0, or by its version ID. That distinction matters in an enterprise setting: you can pin the exact model version production runs on and promote a new one only after it has been tested.

Which tasks and outputs are supported?

According to the data model in the documentation, each prediction can carry a class name, confidence score, bounding box, mask, keypoints and an oriented box. Results can be filtered by confidence threshold or class and exported as YOLO, COCO, CSV or JSON. The table below maps each task to the SDK output and to the usage patterns we see most often in projects; the right-hand column is our commentary.

Task	SDK output	Example enterprise use
Object detection	class, confidence, box	counting parcels at a warehouse gate, checking safety equipment
Classification	class, confidence	sorting product photos into quality grades
Segmentation	mask	measuring surface defect areas, field or roof analysis
OBB	oriented box	tilted parts, vehicles and buildings in aerial imagery
Keypoints	keypoints	verifying assembly points, posture analysis

The documentation also covers batch inference (predict_batch), an async client, GPU preloading to reduce cold starts (warmup), a benchmark helper that measures latency and FPS, downloading weights as ONNX and uploading images to a dataset through the SDK.

How does edge mode work?

The most interesting part of the SDK is edge mode. Once OpenCV support is added with pip install evren-sdk[edge], the EvrenCamera class grabs frames from a webcam, an RTSP stream or a video file, compresses them, sends them over HTTPS to EVREN’s inference cluster and draws the returned result on the image. The documentation describes this as real-time inference on devices without a GPU (Raspberry Pi, laptop, industrial PC).

The defaults are designed to save bandwidth: at most 15 frames per second and JPEG quality 70. The field device does no computation but needs a constant network connection, so what happens on a line where connectivity drops must be answered at design time.

What infrastructure sits behind it?

The architecture drawing in the package documentation shows the SDK connecting over HTTPS/TLS to a FastAPI gateway, which talks gRPC to an NVIDIA Triton inference server running on 8× NVIDIA A6000 GPUs. Users do not manage any infrastructure. In exchange, the documentation is clear that every inference consumes credits: when the balance runs out the SDK raises InsufficientCreditsError (HTTP 402). RateLimitError covers 429, and InferenceError covers 502 and 503.

What does it mean for enterprise software teams?

In computer vision projects the most expensive item is rarely the model itself; it is setting up and running GPU servers and wiring the model into the production application. EVREN SDK simplifies the first two by reducing inference to a single HTTP call. In our assessment, the real work stays in the integration around that call: placing the SDK behind a small middleware service, rather than embedding it in the ERP or manufacturing execution system, lets you manage the key, quota and model version in one place.

If camera frames contain personal data such as faces or licence plates, decide up front, with KVKK (Türkiye’s data protection law) in mind, which frames may leave the site. Predictions can be stored as JSON or COCO and loaded into a reporting database; for auditability, record the model version next to every result.

Pre-pilot checklist

• Have you picked a single camera and a single decision point for the pilot?
• Is the model version pinned with a tag (such as :v1.0)?
• Is it written down what happens to frames during a network outage (skip, buffer, alert locally)?
• Are there retry and queueing rules for 402, 429 and 503 responses?
• Are monitoring and alerts defined for credit balance and error rate?
• Is there a masking or sending rule for frames containing personal data?

EVREN series

This is the first instalment of our series following the EVREN platform step by step. In June 2026 the platform was introduced to universities through YÖK; in September an LLM inference service with 11 models opened, followed by the broad launch of the national AI platform EVREN. For the technical details of wiring the API into enterprise software, see our EVREN API integration guide.

How can Aksiyon Soft help?

Our Samsun-based team works remotely with organisations across Türkiye to connect computer vision services to existing systems, adding planned on-site visits when a project needs them. Through our API and integration service we design the middleware that manages SDK calls, the queueing and retry logic, and the hand-off of results to ERP or reporting. For organisations that want to manage several data flows in one place, our API and data integration platform solution provides a solid foundation.

Our delivery model starts with a discovery session, continues with a single-camera MVP, grows through two-week sprint demos and ends with hypercare and SLA-backed support after go-live. Teams considering the LLM side can start with our article on the AI-ops layer and LLM router.

Frequently asked questions

Is EVREN SDK paid?

The package itself is Apache-2.0 licensed and installable from PyPI. According to the documentation, however, every inference is deducted from your credit balance on the EVREN platform, and credits are earned through the platform’s contribution-based model.

Is model training done with the SDK?

No. The 0.7.0 documentation defines the SDK as a tool for running inference against models trained on the platform. It can upload images to a dataset and download weights as ONNX, but training takes place on the platform.

Does edge mode work without an internet connection?

Not according to the documentation. In edge mode, frames are sent from the device to EVREN’s GPUs; the device only captures the image and draws the result, so a stable network connection is required.

Which Python version is required?

Python 3.10 or later, according to the PyPI listing. The core dependency is httpx; the [edge] extra installs OpenCV for camera and video support.

Is it compatible with our YOLO or COCO based tools?

Results can be exported to YOLO text format, a COCO list, CSV and JSON, so hand-off to existing labelling and reporting tools works without writing extra converters.

Is Aksiyon Soft an official EVREN partner?

No. Aksiyon Soft has no partnership with SSB or EVREN. This article is industry news based on public sources; our integration support is an independent software service.

Let’s talk about your project

If you want to connect camera or image data to your existing systems, leave a short note through our contact form and we will scope the pilot together in the first call.

Sources

• PyPI — evren-sdk project page and release history (0.3.0: 14 Mar 2026)
• PyPI — evren-sdk 0.7.0 documentation (12 Apr 2026)
• EVREN Platform — official site
• Apache Software Foundation — Apache License 2.0