
I arrived at this competition on a high. I had just won an award at a biotechnology hackathon and the challenge couldn’t have suited me better: use artificial intelligence to solve a problem that was already given. The challenge: governing and controlling the use of AI in development teams.
I turned it over, argued with the AI for a good while and ended up at a question. Today it is normal for everyone in a company to work daily with an AI copilot. But who audits those copilots? Who reviews what people do with them? How do you know you aren’t throwing money away by giving the whole staff access, that they are using it well, that nobody is leaking information without meaning to?
That is where Atlas came from: a copilot that audits the copilots.
Atlas is a judge that sees everything. It runs locally, it is lightweight and it sends no data anywhere. It doesn’t intercept or block anything either: it analyses the telemetry those assistants have already left behind, every day or every week, and gives you answers: which departments could make better use of the budget, which are letting keys slip out, where the usage policies aren’t being followed.
There is one decision I defend within the project itself. The calculations, the policies and the key detection aren’t done by the AI, they are done by tools built for that, because AI can get things wrong and in an audit that isn’t acceptable. The language model, which also runs locally, is kept for two things: writing a report people can understand and catching what the rules can’t see, such as a password written in the middle of a sentence. And even there it isn’t left alone: a verifier checks that every figure in the report comes from the data. Use AI when it is really needed, and don’t shoehorn it in.
In the demo, using synthetic data, Atlas reviewed 6,826 uses over eight weeks and showed something uncomfortable: 60% of the spend had gone on answers nobody ended up accepting.
I thought it was a good idea. The kind I would be willing to use in my own company tomorrow.
What was left was selling it. A one-minute pitch.
In the business world you can have the worst idea in the world, and if you sell it better than the person next to you, they will probably pick you. What stays is the first impression, even if someone later looks into it and finds it wasn’t as good as it seemed.
I thought I had both. Public speaking has always come easily to me, and I brought not only a solid idea but a carefully worked pitch. It grabbed attention from the first second, went straight to what mattered without lingering on trivia and closed with a broadcaster’s voice. It didn’t let anyone get bored for a moment.
I listened to everyone else’s pitches. I know it sounds immodest, but I walked out convinced that none came close to mine. Communication is where I stand out most.
Do you know what the result was? I didn’t even make the top three.
This story ends on a bitter note and I’m not going to dress it up. I did things as well as I know how, I was fully convinced I would at least make that top three, and I didn’t.
When something like this happens I like to stop and think about why. Why am I seeing something that others don’t? Because usually, when that happens, the one who is wrong is you and not everyone else. To this day I still don’t know why I wasn’t chosen. But I’m sure I did something wrong.
What I would have been hugely grateful for is someone telling me. And that is the thought I’m left with: in a competition meant for students to learn and develop their skills, why is no feedback given to those who didn’t reach the expected standard? It is the only way they will correct their mistakes next time instead of repeating them.
Technical summary
A prototype I built for the 2026 Atmira × OpenAI × Big Data Hackathon. It reviews exported AI coding-assistant usage to report costs, adoption, possible sensitive-data exposure and compliance with usage policies. It runs locally, using fixed rules for calculations and a local language model for interpretation.