24 January 2026
What does it take to design Chat-GPT-like software that can resolve user tasks without calling external providers? This text was meant to describe my software experiment called RealSpark, a small example on how you can combine off-the-shelf AI models into a self-contained application that aims to debunk AI fakes in art paintings shown online. I realized that the growing importance of AI sovereignty makes the project a good excuse to capture the picture of where we are heading after the AI hype wave turns into the “new normal” of everyday (tech) life.
The task of creating an application that is based on multiple self-hosted AI models might sound like a somewhat challenging one. However, with all the progress in machine learning, its influence on engineering, even a small software shop is now capable of whipping something like that up.
When chasing independence in building with AI, there are at least three technology ingredients that pop up. The first one would be a hardware and geographical deployment of a solution that is a story of its own. System requirements to run a model (inference) are typically lower than building a model (training). Planning for inference may include use-case optimization with a specialised model, and can take into account running on a wide range of platforms from cloud servers to edge devices. While training models have become accessible via global service providers, emerging industry standards like Open Neural Network Exchange (ONNX) are opening doors for independent model training infrastructure without a framework lock-in and deployment fragmentation. Hidden in plain sight, when designing ad-hoc solutions, model evaluation remains one of the most effort consuming components, yet a linchpin of any succesful AI-based service pointed toward chaotic and always-changing real world use cases.
The second ingredient is a buzzling ecosystem of rapidly growing, ready to use, AI models with permissive licensing, such as the open platforms like Hugging Face, Replicate or ModelScope. As for now we can benefit from the tight competitive race between big players or country backed open projects like the Chinese DeepSeek. Once finding access to a good AI model candidate, let's not overestimate community effort to establish interoperability standards of the execution time. Here comes the popular Model Context Protocol (MCP), a way of setting up adapters between AI-based apps and external tools, data, and services. With it, knowledge limitations of your model or its hallucinations can be mitigated by ground truth of curated data sources.
The third ingredient is the AI Agent assisted standardized Software Development Life Cycle loop: plan → analyze → design or specify (including specifying tests) → implement → test → integrate → deploy.
At the end of 2025, AI-assisted development reached a level of maturity that marked a point of no return. Software development is now focused even more on engineering practice. Pure syntax knowledge is becoming less relevant. Setting up configurations, building validation, and accessing knowledge are now quicker than ever. All these bring a new wave of software experiments to come.