> welcome to my corner of the net. scroll down, or talk to my AI.
yan@yandesbiens:~$ ls ~/projects/ --detail
a self-evolving cognitive architecture
An autonomous reasoning system: idle cognition, semantic memory, a "cognitive market" that routes work between dozens of local models, and dream-cycle crystallization. Adaptive across every major subsystem.
train LLMs on one RTX 4090
A full local LLM-training stack — 30M to 500M parameter models, end-to-end, with a web UI. Built to make from-scratch language modeling practical on consumer hardware.
an AI-governed procedural world
A deterministic Python simulation steered by a local LLM "world-spirit", visualized through a Three.js mobile dashboard. The model nudges the world; the world stays reproducible.
a cognitive kernel for agents
A terminal-native agent runtime: a tool bus, pluggable LLM providers (Cloudflare Workers AI, Ollama, OpenAI, Gemini), SQLite memory, and an event-driven loop. Runs anywhere, even Termux.
an agentic loop over ChatGPT-web
A terminal agentic loop that drives ChatGPT through the web — prefix-based conversation threading, a reformulation layer that compiles and steers each prompt, and clean session controls. No API key required.
yan@yandesbiens:~$ cat about.md
I'm Yan Desbiens — I build AI systems that run on my own hardware instead of someone else's cloud. A single RTX 4090, a lot of late nights, and a belief that the most interesting intelligence is the kind you can own.
My work spans LLM training from scratch (ForgeLM), self-evolving cognitive architectures (Hermes/NeuroArch), and AI-governed simulations (AEON). I like determinism, local-first design, and tools that feel like a terminal.
# currently: pushing autonomous cognition further every single day.
yan@yandesbiens:~$ ./chat --talk-to-my-ai
yan@yandesbiens:~$ tail -3 ~/blog/