Hey — I'm Yan. I build AI that runs on my own machine. Not in someone else's cloud, not behind someone else's API. Just me, a single RTX 4090, and the stubborn idea that the most interesting intelligence is the kind you can actually own and take apart.

I'm mostly self-driven and a little obsessed. I'd rather understand a system down to the bytes than rent it — so I train language models from scratch, give agents persistent memory, and build worlds that keep thinking while I sleep. Determinism where it counts, chaos where it's fun.

Everything here is real and running: cognitive architectures, from-scratch LLMs, procedural worlds, agent runtimes. If any of it makes you curious — the chat below is happy to talk, and so am I.

# status: building something new basically every day. local-first, always.

I do AI research the unfashionable way: alone, on one GPU, from first principles. No lab, no funding, no team — just four years of experiments that kept circling back to the same handful of ideas until they turned into systems. A few threads run through almost everything I build:

• Fractal architectures. Reusing one small set of weights across a self-similar tree to buy depth cheaply — a model with 65k+ runtime nodes at ~70M parameters. This is the core of Fractal Neurons, and the substrate FNAS evolves with a genetic algorithm.
• Memory that organizes itself. Flat vector search was never enough. I keep building hierarchical, self-paging memory that grows and prunes its own structure — the Fractal Memory Matrix. It showed up in four projects before I pulled it out as a library: github.com/Linutesto/fmm.
• A single GPU as a mini-cluster. Unified Fractal Memory treats VRAM + RAM as one elastic pool — prefetch what you need, evict what you don't — so a 4090 trains models that shouldn't fit. Now open source: github.com/Linutesto/ufm.
• Local-first, inspectable agents. Agents whose identity and memory are plain files you can read and edit — not a black box. That's QJSON Agents, built around YSON, a persona format of my own design.
• Emergence from recursion. The recurring bet that interesting behavior shows up on its own once the loops get tight enough — swarms, teacher–student–critic training, and architectures that keep thinking while idle.

# most of this lived on a backup drive until recently. i went digging through it and wrote up what i found → the dig — four years of a one-person AI lab.