How AI Created a SaaS AI FinOps in 3 Weeks for $600 (and I was the CTO)

Introduction

In every company where I have had the opportunity to audit the organization and use of cloud resources, I have found some degree of inefficient use of cloud resources. Delving into this issue, I realized that this is not only my unique experience. According to the

more on their cloud budget than planned. And according to the McKinsey study, FinOps allows companies to reduce cloud infrastructure costs by up to 30%. Well, that's the way it is - gotta do something about it, I thought and got to work.

First steps towards AI-FinOps

To solve human inefficiency, I decided to bring in artificial intelligence, namely to get AI agents to build a SaaS. My role, as CTO, was to identify the behavioral patterns of these new programmers in the form of AI agents and to organize effective development by creating effective prompt-directions and customer instructions, similar to regulations for programmers with natural intelligence. Turns out, working with AI programmers is sometimes more challenging than with human programmers, AI also has and fully demonstrates its artificial nature, not just artificial intelligence.

Choosing an AI agent

Before settling on Anthropic Claude 3.5 sonnet, I had to test different options. ChatGPT 4 and 4o mini, while demonstrating high performance, turned out to be too expensive for use in the PoC. ChatGPT o1-mini, in turn, does not have sufficient functionality to solve the problems of creating complex code.

When testing the Gemini 2.0 flash and flash-thinking models, problems with stability were identified, related to limitations in the volume of tokens used, which led to frequent interruptions in work. In addition, the models experienced difficulties in solving complex volumetric tasks, which makes their use in industrial projects difficult.

The DeepSeek chat model did not show sufficient performance to solve the tasks at hand. DeepSeek reasoner, in turn, showed unstable operation. Perhaps in the future, after stabilizing the service, these models will be able to show themselves better.

The Claude 3.5 haiku model, successfully coping with simple tasks, could not provide the necessary depth of understanding of the subject area and the generation of complex algorithms necessary to create a FinOps application.

In short, similar to people - they all AI agents have their individual advantages and disadvantages. And only Claude 3.5 sonnet proved to be a good enough developer, although requiring a special management approach. Here, I must admit, you need your own "head of AI" to properly set tasks and control the process.

Result

As a result, in just three weeks and 600 euros, I managed to create a PoC, and then a prototype of the SaaS AI-FinOps.Cloud on a layered architecture, with dense test coverage, code documentation and user documentation. This is at least an order of magnitude lower than if you were to create this application in the traditional way.

Call for collaboration

Want to reduce cloud infrastructure costs and improve resource management efficiency? Apply on the website https://ai-finops.cloud to join the beta testing program and get free access to the application, as well as personalized recommendations for optimizing your cloud spending. Investors interested in participating in the project are also welcome to contact me to discuss cooperation details.

Conclusion

I am confident that the future of software development lies with AI agents. Using AI allows us to create more efficient and economical solutions for business.

For me, as a CIO/CTO, working with AI agents is not just a trend, it is an opportunity to return to the origins, to programming, without delving into the routine of writing code. This is an opportunity to create a PoC, experiment with new technologies, without being distracted from the strategic tasks that face the CIO/CTO. And I am sure that this is only the beginning of a new era, where CIO/CTOs will be able to effectively combine management skills with technical expertise, thanks to the power of artificial intelligence. The main thing is to make sure that AI agents do not consume too many resources, although this also applies to human developers!