Using AI as a Small DevOps Team

If you ask an AI model to build an entire application in one go, you usually get code that looks fine on the surface, but only half works. The model forgets earlier agreements halfway through, makes loose assumptions about security, and leaves you with something that is painful to maintain.

I noticed this early on when I started experimenting with AI. Today, I take a completely different approach: I do not treat AI as a single clever chatbot, but as a small DevOps team that I manage myself.

Setting up dedicated roles first

When I start a new project, I do not jump straight into prompting or coding. I first think about what roles are actually needed to build the project properly.

I put those role definitions into simple text files (markdown). For example:

• Project manager: keeps an eye on the schedule, the task list, and the big picture.
• Security: only cares about safe storage of passwords, API tokens, and network connections.
• Frontend: focuses on usability and how buttons and layouts behave on mobile and desktop.
• Quality and testing: hunts for edge cases and thinks about ways the app might break or freeze.

The big advantage is focus. A security role looks at code very differently from a role that just wants to finish a user interface as fast as possible. Because the roles are separated, they challenge each other and catch mistakes that I might have missed myself.

Picking the right model for the job

I currently have active subscriptions across several AI providers, including Claude, ChatGPT, and Google Antigravity. Not to keep them all open at once, but because each model has distinct strengths:

• Claude Opus (via Cowork) acts as the team lead. It oversees the whole project, distributes work across the different roles, and logs decisions as we go.
• GPT (Sol and Terra) handles heavy programming work and deep security checks.
• Gemini 3.7 Flash does quick reviews, text summaries, and smaller administrative tasks where I do not need heavy computing power.
• Claude Sonnet is there for quick, lightweight tweaks in between.

Working with Agile and Sprints

Just like a real DevOps team, I work using the Agile methodology. Instead of prompting features at random, I work in short, focused Sprints.

At the beginning of each sprint, I review the backlog and work with the orchestrator (Claude Opus) to define the sprint goal: for example, setting up authentication, the database schema, or a specific user flow. Tasks are then distributed to the specialized models. At the end of the sprint, we test everything end-to-end, log any bugs or edge cases, and record key decisions in a decision log. This keeps the codebase stable, organized, and predictable.

Real-world projects

To see how this works in practice, here are three recent projects:

1. The rental housing search tool

A good friend and his girlfriend urgently needed a rental apartment in the private sector. In today's housing market, refreshing websites by hand gives you almost zero chance. By the time you spot a listing, all viewing slots are usually gone.

So I built a small tool that monitors multiple housing platforms day and night. It checks for new listings every few minutes, filters by maximum rent, floor space, and location, and immediately sends a notification to their phones when a match appears.

My AI team was a huge help here: the security role made sure webhook tokens were stored safely, while the testing role made sure we did not spam them with duplicate alerts when a listing got refreshed. Within a short time, they managed to book several viewings.

2. Syncing an intranet calendar to a smartphone

Two of my girlfriend's colleagues had a common frustration: their work schedules were locked inside an isolated corporate intranet. They had no way to view their appointments on their phones while on the road.

I built a secure backend integration that reads their schedules from the intranet and automatically pushes them to the default calendar app on their smartphones. The role setup worked great here as well: one role built the API integration and handled network timeouts, while the security role verified that no unnecessary personal data was ever cached or exposed.

3. A new consumer platform in development

Alongside these smaller tools, I am currently working on a much bigger project. Together with a business partner, I am quietly building a new consumer marketplace platform for the Dutch market. The idea came from everyday frustration with existing platforms: when looking for second-hand items online, you often have to dig through endless commercial resellers, dropshippers, spam, and paid ads before finding genuine private sellers. We want to bring back a clean, direct, and straightforward peer-to-peer experience.

Because this is a full-scale platform with user accounts, authentication, databases, and strict security requirements, AI orchestration plays a critical role. Different models review each other's code, stress-test database migrations, and keep the overall architecture clean.

The project is currently in the final stages of its alpha phase. It remains strictly work in progress for now, with the goal of launching by the end of this year.

Staying in control of the final result

AI allows you to prototype and build things much faster than before. But the most important rule remains: you always have to check the work yourself.

AI can sound very confident while handing you a solution that is full of logic errors or security flaws. That is why I inspect and test every single line of code and configuration myself before putting it to use. AI speeds up my workflow and helps me stay sharp, but I remain fully responsible for the end result.