I'm Mihai.
I love building things. Guitar amps first, then an internet service provider, wireless networks in Latin America, mobile security platforms and machine translation products.
Now I build AI agents that watch videos for you and surface what matters.
A few of my builds are below. Have a play.
It started with guitar amps.
Before computers, it was music. All through high school I collected guitars and effects.
Then I built my own amp: a Marshall JCM2000, hand-wired from the schematics, tubes and all.
Marshall JCM2000 copy · tubes · hand-wired
Then I built an internet service provider.
All through university there was no broadband where I lived, only dial-up. So we built a network in our neighbourhood, and it spread across the city.
Cisco routers were out of reach, so the network ran on cheap PCs with Slackware Linux, quick to swap when one died or got struck by lightning. We patched the Linux kernel to act as a layer-7 router, so the network shared traffic fairly between our 2,000 customers.
The BBC later made a documentary about the teenagers who wired Romania. I was one of them.
2,000 customers · Slackware · layer-7 routing
From wires to wireless.
I finished uni at UPC in Barcelona, measuring quality of service across European IP networks.
Then came WiMAX at Alcatel-Lucent, now Nokia: mobile broadband over the air. I tuned its quality of service and took it to operators in the Dominican Republic, Venezuela and Brazil.
WiMAX · technology introduction · Latin America
Finding malware with machine learning.
From mobile networks to mobile phones. Back then everyone thought their phone was safe, and Lookout set out to keep it that way. The malware fought back by rewriting itself, so no two copies looked alike.
So machine learning took the first pass. Binary similarity linked every copy back to its family, and the researchers focused only on what was truly new.
And as the phones became sensors, we started protecting businesses and their mobile fleets.
polymorphic malware · binary similarity · machine learning · cybersecurity
Machines that translate, and know what they got wrong.
As general manager of Language Weaver, I took the latest machine translation research and productised it for enterprises, governments and the public sector, on-premise or in a secure cloud. The focus was adaptability: we adapted the models to each customer's content.
In the end, the interesting question wasn't “is this translation good?” but “which parts aren't?” Quality estimation scores every segment, and only the ones that fall short go on to a large language model to be revised. I co-invented that automatic post-editing, and we shipped it as Evolve.
Beyond the product, Language Weaver's AI went to work inside a large, publicly listed language service provider, automating processes across its language operations and lifting its margins.
SMT → RNN → LSTM → Transformer → LLM · patents in QE and content routing
Now I build agents that watch videos for you.
From text to video, and back to building something of my own. You can't read a video like a book: no highlighting, no notes, no skimming. So I launched Video Highlight as a Kindle for video: highlight the transcript, take notes as you watch, get timestamped summaries.
Now agents can do the watching, and the whole idea finally works. Give one a two-hour video and it watches every minute, then comes back with the moments that matter, screenshots already taken.
I'm writing code again, with a few agents working beside me.
350,000 users · two hours in, five moments out
And something new is on the way.
A secret project, launching soon. It listens.