Alexis

I remember the first week I walked into the Apple campus in San Francisco, March 2022. I had just finished a Master's in Machine Learning at Georgia Tech, done stints at GoPro and HERE, shipped a platform at CES in Las Vegas. I thought I understood scale. I did not. At Apple, "a billion users" is not a marketing phrase. It is the actual constraint inside every design meeting. Every model you train, every feature you architect, every latency decision you make has to hold at a scale that is genuinely hard to conceptualize. As Lead AI and ML Engineer Architect on Apple Intelligence, I was building systems that had to work for a retired teacher in rural Iowa and a teenage developer in Seoul and a grandmother in Lyon who has never opened a terminal in her life. Same system. Same model. One shot. And here is what that experience taught me that I did not expect: at that scale, the hardest problem is not technical. It is human. We were building features that touched how people communicate, how they find information, how they remember things. The engineering bar was extraordinary. But what actually broke features in production was never the math. It was the moment the system stopped feeling like it knew the person using it. The moment it felt generic. That is when users left. Not dramatically. Quietly. The way you stop using a tool that never quite fits your hand. I spent nearly six years at Apple watching this pattern repeat. A feature would be technically impeccable and emotionally flat. Another would be rougher, less polished, and people would love it because somewhere in the design someone had made a choice that felt personal. That felt like the product had noticed them. --- There is a specific frustration that builds when you are good at something and you can see exactly where the gap is. By 2021 I was building ML systems I was genuinely proud of. Point of view classifiers running at 90-plus percent accuracy. Federated learning models that kept user data private while still improving. Real, hard, impactful work. But every AI tool I used outside of work reset every single time I closed the tab. Claude did not know me. ChatGPT did not know me. Gemini did not know me. I would explain my context, my goals, my communication style, my history, and then come back the next morning and do it again from zero. Like hiring a brilliant consultant who has amnesia between every session. I kept thinking: this is absurd. We have spent decades learning that the best human relationships deepen over time. They accumulate context. They adapt. A great mentor does not ask you the same questions every meeting. A great colleague learns how you think. The relationship compounds. And yet every AI tool we were building was architecturally incapable of doing that. Not because it was hard. Because nobody had made it the core design principle. That question — "what if the AI actually knew you?" — became the seed of everything I built next. --- I need to say something about music here, because it is not a side note. It is actually central to this. I make deep house and techno music. I produce in Ableton, usually late at night after a long day of debugging deployments and writing architecture docs. The project I am most attached to I called Deeper Heights, which is about the harmony between deeper sounds and higher sounds, the way the low frequencies carry the weight while the higher ones find the light. It sounds like a metaphor for engineering because it is. When I perform, I perform as Cherry Set. I played Miami Music Week in 2023. I curated events in Paris, hosted a night near the Golden Gate Bridge in San Francisco. And what I learned on those floors, watching how music makes people feel seen in a room full of strangers, is that the feeling of being understood is not a soft, optional feature. It is the whole point. It is what separates a set that people remember from one they politely endure. Technology that does not make you feel seen is music with no bass. Technically correct. Emotionally hollow. --- So I built Uare.ai. And the architecture is not complicated to describe, even if it is genuinely hard to build: instead of you adapting to the AI, the AI adapts to you. Everywhere. In every tool. On every device. We built a portable personal AI that carries a compressed representation of who you are into every interaction. Your voice. Your values. Your communication style. Your current goals. Your relationship history. The way you make decisions. The stories that shaped you. Not as a profile you fill out, but as a living graph that learns from how you actually work. The insight from the Uare.ai architecture is this: the experience should feel like the AI just knows you, not like it's reading from a file about you. That one principle drove every technical decision. The Graph-RAG knowledge graphs. The adaptive voice systems. The RLAIF loops that keep improving over time. The browser extension that activates your personal AI