scroll ↓

…

Hi, I'm Leonel Salvo, aka Pinguinista.

I love penguins, pixels and creativity in all its forms.

Below: too many of my areas of interest, where I burn my neurons most days.

Pinguinista · tools

Involut

Understand what you build. Grow through the work.

Your work with AI can also be your practice. Involut applies Auctor to code changes: decisions, questions, evidence and room to intervene.

Explore Involut In development · method in draft

Pinguinista · tools

Augmented Mastery

A philosophy for the purpose, a method for the practice, and a product to bring it into your work. Each has its own role.

Read the manifesto · PDFExplore the Auctor method

Augmented Mastery

The philosophy. Delegate execution while developing understanding, judgment and the ability to intervene in your work.

Auctor

The method. Get oriented, take a position, check evidence and revisit what you learned. A draft whose outcomes still need to be tested.

Involut

The product. A first application to code, with cards, questions, original material and notes to work through with your agent.

Interests · experiments

Neural networks from first principles.

PyTorch, one self-checking component at a time, on a single RTX 3090. No nn.Transformer, no wrapper I can't explain. Next down the stack: CUDA and Triton kernels.

A token moves through attention and feed-forward layers. Four blocks shown to make the structure readable.
PredictionScores the next token
SwiGLUTransforms each token
GQA attentionConnects token context
RMSNormNormalizes each layer
EmbeddingTokens → vectors
DenseArchitecture study · simplified
A compressed attention state, a shared expert and two routed experts per token. Eight experts shown as a schematic.
PredictionScores the next token
Router + expertsSelects paths per token
MLA attentionCompresses the KV cache
RMSNormNormalizes each layer
EmbeddingTokens → vectors
MoE + MLAArchitecture study · simplified
Two prediction heads, for t+1 and t+2. BPE starts with bytes; Muon changes the optimizer, not the forward-pass architecture.
Multi-token headsPredicts t+1 and t+2
SwiGLUTransforms each token
GQA attentionConnects token context
RMSNormNormalizes each layer
EmbeddingTokens → vectors
MTP + BPEArchitecture study · simplified
  • A token moves through attention and feed-forward layers. Four blocks shown to make the structure readable.
    modern-nanoGPT

    The dense transformer rebuilt in PyTorch: RMSNorm, RoPE, SwiGLU, GQA and tied weights. Each component checks itself.

    ↳ A token moves through attention and feed-forward layers. Four blocks shown to make the structure readable.1 / 3
  • A compressed attention state, a shared expert and two routed experts per token. Eight experts shown as a schematic.
    nano-moe-mla

    Sparse MoE + MLA in one model, with a router-specialization probe and a seed-averaged ablation.

    ↳ A compressed attention state, a shared expert and two routed experts per token. Eight experts shown as a schematic.2 / 3
  • Two prediction heads, for t+1 and t+2. BPE starts with bytes; Muon changes the optimizer, not the forward-pass architecture.
    frontier-llm-techniques-2026-Q1

    Muon, Multi-Token Prediction and a base BPE tokenizer: self-checking implementations, plus an AdamW vs. Muon benchmark.

    ↳ Two prediction heads, for t+1 and t+2. BPE starts with bytes; Muon changes the optimizer, not the forward-pass architecture.3 / 3
RMSNormAttentionFeed-forward / experts

Career

Head of AI Engineering

  1. 2018

    A degree in electronics

    Hardware first: circuits, signals, and the habit of taking a thing apart to see how it works.

  2. 2020

    Solidity full-stack developer

    Into web3: contracts on one side, the interfaces that talk to them on the other.

  3. 2021

    CTO · Astronaut Penguin

    My own software factory. Work for the TON blockchain and a string of web3 projects, from the architecture down to the deploy.

  4. 2022

    Senior frontend engineer

    B2B travel: flight and hotel packages, sold to agencies rather than to travellers.

  5. 2023

    Lead frontend engineer

    Decentralised information and private AI — the first time the two halves met.

  6. 2024

    Head of Product Discovery

    Deciding what to build before building it, and being answerable for the choice.

  7. 2025

    Head of AI Engineering

    A private, multi-tenant enterprise AI platform: composable agents and retrieval over a customer's own documents, in production. Architecture kept private.

Seven years from soldering irons to agents, and the same method throughout: build it myself first, then trust the tool.

What I have built with

Web3
  • Solidity
  • Hardhat
  • FunC
  • wagmi
  • web3.js
  • ethers.js
  • paper.xyz
Web
  • TypeScript
  • React
  • Vite
  • Next.js
  • NestJS
  • FastAPI
  • PostgreSQL
  • Redis
  • Docker
  • Kubernetes
  • Azure
AI
  • PyTorch
  • Python
  • Transformers
  • RAG
  • Triton
  • CUDA
LinkedIn

Say hello.

agustin.leonel.salvo@hotmail.comLinkedInGitHub

© 2026 Leonel Salvo · Pinguinista

All rights reserved. The original writing, visual assets and this site are not licensed for reuse. Code I open is published in its own repository, under the licence stated there.