Lead AI Developer — production AI infrastructure by day, neural networks rebuilt from scratch by night, so I actually understand what I ship.

Across AI engineering, games and tools, one question: can I explain and intervene in what I build?

• 🔬  Research — I rebuild modern architectures from first principles: no nn.Transformer, no wrapper I can't explain.
• 🤖  Work — I lead a private, multi-tenant AI platform: composable agents + RAG over a customer's own documents, in production.
• 🎮  Games — a one-person Godot studio, Pinguinista: Una última cosa in development; Records of the Truth Bureau on hold. Neither released.
• 🧰  Tools — Involut, applying the Auctor method to learn from work produced with coding agents.
• 📚  Permanent student: a degree in hardware / electronics, and pushing from LLMs toward CUDA and quantum.

The modern stack rebuilt one self-checking component at a time, in PyTorch, on a single RTX 3090.

modern-nanoGPT dense transformer · RMSNorm · RoPE · SwiGLU · GQA	The modern dense transformer. The GPT-2 skeleton upgraded to the 2026 open-weight stack — RMSNorm · RoPE · SwiGLU · GQA · tied weights — built one self-checking component at a time. Trained on a single RTX 3090 (val ≈ 1.48).
nano-moe-mla sparse MoE + MLA · routing probe · ablation	The sparse template: MoE + MLA in one model. Built from scratch with the instruments to study it — a labeled multi-domain corpus, a router-specialization probe (mutual information), and a seed-averaged stack ablation that confirms the balancing ↔ specialization tradeoff.
frontier-llm-techniques-2026-Q1 Muon · Multi-Token Prediction · BPE	Frontier techniques, from scratch. The Muon optimizer (used to train Kimi K2) and Multi-Token Prediction (DeepSeek-V3, Gemma 4, GLM, Qwen), plus a base BPE tokenizer — each self-contained and self-checking, with an AdamW-vs-Muon benchmark.

I lead a private, multi-tenant enterprise AI platform — composable agents and retrieval over a customer's own documents, shipped to production. (Architecture kept private.)

Augmented Mastery is the philosophy: delegating execution can coexist with developing understanding, judgment and the ability to intervene in your work.

Auctor is the method under development: get oriented, take a position, check the evidence and revisit what you learned. Agreement with an agent is not proof of mastery. The method is a draft; its learning outcomes still need evaluation.

Involut is the product applying it to code: decision cards, questions before revealing, access to original material and notes for the next exchange with your agent. TypeScript · React · three.js. The product repository remains private during development; no public demo is claimed.

Explore the philosophy, method and product · Read the manifesto · Document repository

_{Manifesto and Auctor texts: CC BY-SA 4.0. Involut software: AGPLv3. Separate projects, with separate licensing scopes.}

I didn't invent these — I learned them by rebuilding and verifying each one in code, following Karpathy's Zero to Hero and the open-model papers.

_{Method: from-scratch first; a wrapper is only allowed once I can name what it wraps and the trade-off.}

AI / ML  

Backend  

Frontend  

Infra & also