San Fernandes, AI Engineer with a background in data

01Intro

Language mastery

Somewhere along the way, I mastered every programming language.

Not quite. I've mastered Python and SQL. I've shipped TypeScript, C# and C++, which is different, and most of the difference is knowing when the AI is bluffing.

Before the highly scientific evidence, a little about me.

02About

Who's being evaluated

I'm San, an engineer with a background in data who builds with AI.

Most of my work starts with data: pipelines, warehouses, streaming services, and the retrieval systems behind AI tools. But I've never been good at staying in one lane. If something would be useful, I'd rather build it than wait for it. That's how I've ended up shipping iPhone apps, macOS apps, Windows utilities and OBS plugins alongside the day job.

I see AI as leverage, not a shortcut. It gets me to a working first version quickly, which leaves more time for the parts that need real thought. And honestly, the pace still amazes me. Things that felt out of reach a year ago are now an afternoon's work.

I taught myself to code while working in financial services, where I got curious about the systems behind everyday work: how information moves, where it gets stuck, and what better tooling could change. The Northcoders Data Engineering bootcamp came later, to add structure and depth to what I'd already been building, and I haven't stopped since.

03Benchmarks

Independent evaluation results

In the interest of transparency, I put myself through the same evaluations used to rank the world's leading language models. The results are in, and I've decided to publish them.

san-fernandes / eval-resultsNov 2026

eval --subject san --suites all

5 suites complete. Peer reviewers: 0.

note: subject is model-agnostic

GPQA Diamond

PhD-level science, under pressure

96.0%

↳ GPT-6 Astra got the same. Small world.

OSWorld 2.1

Using a computer without calling IT

89.3%

↳ level with Claude Fable 5.2, purely by coincidence

Humanity's Last Exam

The hardest questions we've got, with tools

74.2%

↳ matches Claude Fable 5.2. I also used tools.

Chartography

Reading charts, no tools

71.0%

↳ identical to GPT-6 Astra. Neither of us squinted.

Terminal-Bench 4.0

Getting real work done in a terminal

66.4%

↳ the Claude Opus 5.5 number, to the decimal

Methodology: self-administered. Scores bear an uncanny resemblance to the current frontier-model leaderboard. Peer reviewed by zero peers. No San was fine-tuned in the making of this chart. Every model on this list has, at some point, worked for me.
04Work

Ground truth

The benchmarks were borrowed. These are mine: data pipelines, a RAG system you can walk through in 3D, native Windows tools, and a working studio.

Repo · 2026GitHub

RAG Galaxy (opens in a new tab)

Retrieval-augmented generation, as a 3D film

Ask a question and follow it through a real RAG pipeline: embeddings, vector search, ranking, and a cited answer from Claude. It's scored against a golden question set and deployed to Azure with Terraform.

PythonFastAPIpgvectorClaudeReact Three FiberTerraformAzure
Studiolittlelotus.co

Little Lotus (opens in a new tab)

My web design & AI automation studio

Shopify stores, fast marketing sites and AI workflows, shipped for real clients.

Web DesignShopifySEOAI Automation
Repo · 2025GitHub

Clip Atlas (opens in a new tab)

Because filing clips by hand is a crime

A native C++ plugin for OBS Studio. It spots the game you're capturing, runs the replay buffer for you, and files every replay, recording and screenshot into the right folder. It started life as a Python script.

C++OBS StudioWindowsAutomation
Repo · 2026GitHub

Display Switch (opens in a new tab)

Windows HDR, minus the settings menu

A tray app that turns HDR on when a game starts and off when it closes. It hears about Steam launches before the game does, and idles at 0.2% of one CPU core.

C#WinFormsWin32Windows
Repo · 2025GitHub

Guardian Content Stream (opens in a new tab)

Event-driven ingestion microservice

Pulls articles from the Guardian's Open Platform API, validates every record, and publishes them to Amazon Kinesis as JSON events. Serverless, and deployed with AWS SAM.

PythonAWS LambdaKinesisPydanticboto3SAM
Repo · 2026GitHub

Music Stream Warehouse (opens in a new tab)

An analytics warehouse, built one layer at a time

An end-to-end data platform for a mock music-streaming service. Raw events land in BigQuery, dbt models the marts, and Airflow runs the schedule.

BigQuerydbtAirflowSQLPython
05Contact

Need someone who ships at the speed of a frontier model?

I'm open to data and AI engineering roles. Say hello. Unlike the models above, I'll remember the conversation.