Yosberto Baro Carbonell · software engineer · Miami

Building to understand.

I build backend systems and small experiments because I like knowing what is happening underneath. I’m especially interested in databases, infrastructure, distributed systems, graphics, and languages that let the type system do more of the thinking.

Currently completing a Master’s in Computer Science and Engineering at the University of Bologna.

See selected work GitHub Email me

A cyanotype print: scattered pale specks at the bottom resolve into a fern-like graph, a grid of halftone dots and a bundle of strands, which feed into an open blueprint framework at the top. — **Plate 0** Observations → patterns → structure.

Field note · current work

Boundaries that survive the source.

At work, one of the systems I’m involved with moves data out of a legacy Oracle system and into the MongoDB system replacing it. The synchronization itself is not the interesting part. The boundary around it is.

Repositories tied to each source produce a shared internal representation. RabbitMQ separates acquisition from downstream processing. Everything after that works against stable contracts, so a source can be replaced without teaching every consumer where its data came from.

Fig. 1 Schematic, not a system diagram: the parts named in the note, in the order data passes them.

Why I keep rebuilding things

It started with taking toys apart.

When I was a kid in Havana, anything that looked remotely high-tech made me wonder what was inside. RC toys suffered the most. I would open them, usually break them, and try to work backwards from whatever I found.

I had no internet and no idea which books I should even be looking for. When I eventually discovered the internet, the problem flipped: suddenly the only real limit seemed to be how quickly I could get all that information into my head. The instinct never really went away. I just got better tools—and, eventually, fewer broken toys.

Understanding is when the last gap in the web closes. A concept from chapter two turns out to generalize one from chapter eleven; one rule becomes a consequence of two others; the jump stops feeling like a jump.

A working definition of “understand”

Selected projects · three plates

Things I built because I had questions.

The project usually starts with something I can use but cannot quite explain yet.

A cyanotype print of a prerequisite graph drawn like a pressed fern: solid pale nodes below, faint outlines above, and one brighter ringed node at the frontier. — **Plate I** Prerequisite topology; the ringed node is what to study next.

Specimen 01 · principal studyPython · PostgreSQL

Hermes Curriculum

A curriculum engine that decides what to study next, leaving an AI tutor to teach rather than plan.

Why: It bothered me that LLMs could retrieve so much information while I could only absorb it so quickly. I wanted more than summaries: a system for turning course material into an actual path through the subject.
Learned: How graph structure, FSRS scheduling, verification gates, and hexagonal boundaries can become one usable learning system.
Touches: Python · PostgreSQL · pgvector · MCP · knowledge graphs

The question underneath it: how can human reasoning and an LLM’s ability to retrieve information help each other without pretending they are the same thing? I don’t have a definitive answer. Hermes is how I explore it.

I use it for my actual university exams. It helped me prepare for Cybersecurity, completed with 29/30.

Read project notes →Source ↗

Specimen 02 · rendering study

Zero Renderer

A route into Rust and graphics programming that began with raw rendering math before moving to wgpu.

I built it because using a game engine before understanding the pipeline felt like skipping the interesting part. The project moved from CPU calculations to shaders, GPU resources, and the mechanics behind drawing a pixel.

Rust · wgpu · shaders · graphics pipelines

Read project notes →Source ↗

A cyanotype print of a wireframe icosahedron whose projection rays pass through a picture plane and clear a silhouette of pixels in a coarse grid. — **Plate II** Projection to raster: geometry through a picture plane, into pixels.

A cyanotype print of a stream of packets on a single strand flowing into a grid of memory slots, some filled pale, some still blue. — **Plate III** A protocol stream entering memory.

Specimen 03 · protocol study

Redis Clone

A small Redis implementation in Go, built to understand RESP, in-memory data structures, TCP servers, and concurrent connections from the inside.

Go · TCP · RESP · concurrency

Notes →Source ↗

More experiments

Full project archive →

From the notebook · July 2026

What I’m following now.

Constraint before generation

Give the model a cage it can measure itself against.

I don’t trust a model with a whole end-to-end project and no external standard. On a recent legal-editor replacement, I built the infrastructure first: the useful measure is how many items match the old system byte for byte.

The evaluator penalizes regex-heavy solutions. I would rather fix the structure of the problem than hide it under string hacking and call the result done.

Current question

How should I reason about concurrency before reaching for an abstraction?

Studying Concurrent and Parallel Programming: synchronization, memory models, scheduling, and the difference between code that runs and a model I can trust.

On the desk

Rust Atomics and LocksMara Bos
AI EngineeringChip Huyen
Programming Massively Parallel ProcessorsGPU study curriculum

Language margin

At work, mostly C#. For fun, Rust—and right now, Haskell. I keep gravitating toward languages that let the type system carry more of the reasoning.

Interesting problem? Tell me about it.

I’m always looking for the most interesting challenges I can find, preferably with interesting technology behind them. I’m open to work that lines up with those interests and still leaves room to learn.

[email protected] Request résumé