No Python. No GPU. Just Go: Building Sigma-Vanguard

01 April, 20256

Most modern AI systems are built around a familiar assumption: Python is part of the runtime. From model loading to inference, it’s almost always somewhere in the stack. While this works well, it also introduces complexity — additional processes, dependency management, and overhead that becomes noticeable in production systems.

Sigma-Vanguard began with a different idea. Instead of following the usual pattern, the goal was to explore whether a complete AI-powered system could run entirely in Go, without relying on Python once deployed. The result is a system that performs real-time identification of military equipment from images, combining object detection with fine-grained recognition in a clean and efficient architecture.

At a high level, Sigma-Vanguard takes an image and determines not only where objects are located, but also what they actually are. Rather than stopping at generic labels such as “vehicle” or “aircraft,” the system is designed to recognize specific types of equipment. This distinction is important in many real-world scenarios, where understanding the exact type of object carries far more value than simply detecting its presence.

To achieve this, the system is structured in two stages. First, an object detection model scans the image and identifies regions of interest. These regions are then passed into a second stage responsible for identification. Instead of relying on a single approach, Sigma-Vanguard evaluates each detected object using two complementary methods.

The first method compares the object against a collection of known examples using vector similarity. Each image is transformed into a numerical representation, and the system searches for the closest match. This approach allows new equipment types to be introduced easily by adding a few reference images, without requiring any retraining. Alongside this, a compact classification model provides more precise predictions for categories it has already learned, offering higher confidence when dealing with known classes. Together, these two methods create a balance between flexibility and accuracy.

What makes this system particularly interesting is how it is executed. All models are exported to ONNX format and run through the ONNX Runtime C library, accessed directly from Go. This means that once the service starts, the models are loaded into memory and remain available for all incoming requests. There are no external processes, no interpreter overhead, and no need to repeatedly initialize models. The result is a system that behaves more like a traditional backend service, while still performing complex AI inference.

This design has practical advantages. It keeps latency low, even when running entirely on CPU, and simplifies deployment by avoiding heavyweight dependencies. At the same time, it maintains enough flexibility to adapt to new data without requiring constant retraining or infrastructure changes.

Another interesting outcome is how effectively the system works with limited data. By leveraging modern embedding techniques, Sigma-Vanguard can distinguish between visually similar equipment types using only a small number of examples. This significantly reduces the effort typically required to build specialized recognition systems, where large datasets are often considered a necessity.

In the end, Sigma-Vanguard demonstrates that building an AI system does not have to mean accepting complexity as a given. With the right combination of tools and architecture, it is possible to create something that is both powerful and straightforward to operate. By removing unnecessary layers and focusing on a clean execution model, the system achieves a balance between performance, flexibility, and simplicity.