The engineer teaching your Mac to read a scam before you do

One of Nicolas Ha's first real users found him on Reddit, where the man had written that spam was wrecking his business, not just his social life. He was careful about what he installed. He refused the beta and asked who had independently verified the security. Nicolas had to tell him the truth: nobody had, and he could not yet afford to pay for an audit. The user only relented once Apple's App Store review had passed, and even then he tested Klar in a separate mail client first.

That caution turned out to be the most useful thing that could have happened. It is also the cleanest way to understand what Nicolas is building, and why he is building it the way he is.

A messaging problem and a privacy problem

Klar is an AI spam filter that classifies your mail on your own device instead of on a server. It ships as a free Apple Mail extension on the Mac App Store, an iPhone app for texts now in beta, and the same engine as a self-hosted milter for companies that run their own mail servers.

The problem it attacks is one most people have simply learned to live with. Spam filters, Nicolas argues, tend to be either good and cloud-based, or private and bad at the job. Gmail's filter is good, but Google reads your email to get there. Apple Mail's built-in filter does not read your email, and is worse at deciding what is spam. That trade-off was tolerable until AI made it dangerous. Scams are now written by machines, they read like genuine mail, and they walk straight through the old rules.

His fix is to put a small AI model on the user's own Mac, one that reads what a message actually means and decides what deserves to reach you. Nothing is shipped to a server to make that call.

The classification runs on the device. Private by design, multilingual by construction, and it costs nothing to operate because it runs on your own hardware.

That last point matters more than it sounds. Because the work happens on the user's machine, there is no cloud bill, no need to raise money to cover inference, and less energy spent per message. The model is also multilingual from the ground up, not an English filter with other languages bolted on afterwards, a detail Nicolas says the industry routinely neglects.

The hard part of someone else's product

Nicolas did not arrive here by accident. He spent ten years in London building messaging infrastructure at scale and security inside regulated environments, always, as he puts it, as the hard part of someone else's product. Spam is a messaging problem and privacy is a security problem, so Klar sits squarely on both subjects he has spent a career inside.

He came back to France, reconnected with the ecosystem there, and noticed email getting worse in a specific way: the machines that now write our messages had also learned to defend themselves against scams, and ordinary inboxes had not kept up. He became convinced there was a better line of defence, and a longer game underneath it, one where AI models are trained for narrow tasks and become extremely cheap to run for exactly those tasks.

The three weeks nothing was being read

The Reddit user, once convinced, ran Klar for three weeks. Then he came back and said he was not seeing any difference, and asked whether he had installed it wrong.

He had not. At some point in those three weeks the Mail extension had quietly unregistered itself from Apple Mail. Nothing looked broken. The app still opened. It still said it was on. And for three weeks a careful, security-conscious user had believed he was protected while not a single message was being read.

He had spent three weeks believing he was protected while nothing was being read. User feedback is gold, and I have other examples of these embarrassing little moments.

Nicolas does not tell that story to flatter himself. He tells it because it is the kind of failure that only surfaces when a real person trusts your product with something they actually care about, and it is the reason he now treats early users less like customers and more like a warning system.

That instinct also shapes the thing Klar most has to earn, which is trust. Rather than ask people to take a security tool on faith, Nicolas made it checkable. The engine is open source under the AGPL licence, and the Mac app ships with no network entitlement at all, so it cannot upload your mail even if it wanted to. Both are things anyone can verify for themselves.

I could not afford an audit, so I did the next best thing: the engine is open source and the app has no network access. You do not have to take my word for it.

What he is betting on

For all the engineering, Nicolas is candid that the hardest part is not the model. It is sales. Ten years as an engineer, he says, means he is good at the part nobody sees and untrained at the part everybody sees. Building in public, asking strangers to trust an unaudited security tool, standing behind the product in his own name: all of it sits outside his comfort zone, and he does it anyway because the end result of making something genuinely useful is worth it to him.

There is also, now, an honest signal of where Klar stands. Of the people who land on its App Store page, roughly one in three go on to install it. The listing converts. The work now is getting it in front of more of them.

Where he wants to take it is unglamorous and specific. Reach a financially independent footing. Hire the first people who are good at the things he is not. Then train models for other narrow tasks, perhaps circling back to fintech and fraud detection, the territory where reading intent and catching a lie is the whole job. For now, the bet is smaller and sharper: that privacy and good filtering were never actually opposed, and that the machine in front of you is powerful enough to prove it.

This profile is drawn from Nicolas Ha's own account of building Klar. Klar is an early-stage venture and the product details are as described by the founder.