I'm Meelis Sootalu, founder of OEJ OÜ. I build web applications and tools with AI, and configure AI assistants. Here I share my projects, practical guides and what I learned while making them.
OEJ started with a practical question: how much real work can you hand to AI when the system around it is built properly? I'm interested in tools that help with research, writing and preparing tasks. How to check the result matters just as much.
I run my own AI agent for research, analysis, writing and directing other agents. I configured existing agent software rather than building a new language model.
It has memory files, a knowledge base, tools and scheduled tasks. Saved instructions and lessons can be reused, but that does not guarantee that the agent retrieves every detail when needed. The case also covers failures discovered during use.
Read the project story →For the 2026 World Cup I built a system connecting match research, statistical modelling, prediction submissions and checks against the results.
The experiment taught me that a prepared prediction is not necessarily a prediction submitted on time. Source checks, submission steps and human approval needed work. I do not claim an unverified accuracy percentage or final ranking.
Read the project story →The assistant searches documentation and drafts a reply in the customer's language. A person checks the content, makes corrections and sends the answer.
A local filter replaces detected personal details before the model request is sent. If a follow-up check finds a remaining identifier, the request is stopped. The filter does not detect every personal detail or guarantee anonymity; human review is still needed.
The standalone PII Guard is a smaller local tool. It does not search documentation, generate replies or send text to an AI model.
Read the project story →The application helps someone describe their circumstances and get an overview of possible benefits. AI interprets the text; separate code handles the calculations. The app does not submit applications or make official eligibility decisions.
I directed AI agents through research, specifications, implementation, testing and release checks. The results view, payments and access controls also had to work together. A chat window was only the beginning.
Read the project story →AI can sound completely sure and be wrong. I want to see where a number came from and how it was checked. If the result cannot be verified, that should be stated.
AI researches, analyses and prepares. Sending, publishing, moving money and changing access require deliberate human approval. A rule in an instruction is not a substitute for a technical safeguard.
When a system handles sensitive data, money or important decisions, first agree what it must not do and where the data may go. Then build the capabilities.
I've seen green tests while the product was still broken. A well-written wrong answer can look right at first. I also check what the user actually sees and receives.
When something goes wrong, I ask which instruction or check could have helped catch it. Those lessons can become reusable working methods. Finding the mistake does not do that work automatically.
I've worked with software, data and technical operations for more than ten years. AI changed how large a system one person can build and keep running. One thing did not change: a bad process with AI on top is just a faster bad process.
If you have a recurring task you want to make easier, get in touch. I start with a specific task, not a promise to automate the whole business at once.
Get in touch