I’m Matheus. I work on machine learning and software, and most of my code is in Python, TypeScript, React and Rust. Each planet here is one of my projects, orbiting a black hole. Click one to land on it.
I’m Matheus Teixeira, from Santos, Brazil. I work on machine learning and software. My research is in quantum machine learning and quantum error correction, and most of my code is in Python, TypeScript, React and Rust.
I started in 2016 as a web developer and IT support assistant at Orion Benefícios. I have studied at FEI since 2019, first a B.Sc. in Computer Science and then an M.Sc. in Electrical Engineering in AI. My dissertation is a hybrid quantum-classical pipeline for breast cancer histopathology, with the data split by patient so that no patient appears in both training and test sets. I am now doing an MBA in Data Science, AI and Analytics at USP on a full scholarship, and I am enrolled in the Ph.D. at FEI on quantum error correction, with a CAPES scholarship and classes from February 2027.
Below are five of my projects, one per planet.
matheus@orbit:~$ whoami Matheus Teixeira: machine learning engineer, software engineer, quantum computing researcher matheus@orbit:~$ cat now.txt MBA in Data Science, AI and Analytics at USP, in progress Ph.D. in Electrical Engineering at FEI on quantum error correction, enrolled, classes from February 2027 Open to machine learning and software roles matheus@orbit:~$ ls ~/shipped isaac3d.com/ amadeosheets.com/ liderplustelemedicina.com.br/ brectus/ allwinner/ quantum-ml-healthcare/ matheus@orbit:~$ echo $LANG pt_BR (native) en_US (C2, EF SET 92/100) matheus@orbit:~$
One project per planet. Land on a planet for the full story.
Interactive physics simulations
Interactive physics simulations for students, with live graphs and equations.
Audio to sheet music
Turns audio recordings into editable sheet music, all in the browser.
Desktop app for AI coding agents
Desktop app that sends requests to a separate AI coding agent for each project.
Orion Benefícios, LÍDER Plus Telemedicina and an esthetics studio
Company websites for Orion Benefícios and LÍDER Plus Telemedicina, a website for an esthetics studio, and an employee portal.
Backtesting trading strategies
Backtests trading strategies on out-of-sample data, with fees and slippage.
My path so far and what comes next. Time runs to the right, and the line goes up for the year in Zurich. Dashed parts are planned.
Final project: a variational quantum classifier for the Iris dataset, compared with a classical SVM.
Full scholarshipProject code ↗Dissertation on quantum machine learning for breast cancer histopathology: 80.0% accuracy and 86.0% F1, with no patient in both the training and the test set.
CAPES scholarshipDissertation code ↗Statistics, Python, data analysis, machine learning, artificial intelligence and analytics.
Full scholarshipQuantum error correction and machine learning decoders. A sandwich year at ETH Zurich is planned for 2028–2029.
CAPES scholarshipBreast cancer histopathology classification on BreaKHis with ResNet-50, patient-wise splits and SVM and QSVM models; experiments on IBM Quantum hardware on how noise affects quantum kernels.
Help sessions and exercise lists on matrices, linear systems, vectors, lines and planes.
Built and maintained the company website and its database-backed systems, and supported computers, networks and infrastructure.
With Reinaldo A. C. Bianchi.
Read the paper ↗Interactive physics simulations
Interactive physics simulations for students, with live graphs and equations. Physics core in Rust compiled to WebAssembly; interface in React, TypeScript and Three.js; physics validated with cargo test, unit tests in Vitest and end-to-end tests in Playwright.
The equations are integrated in a Rust core compiled to WebAssembly. TypeScript runs the scene, and React and Three.js draw the controls, graphs and 3D view.
Each simulation is tested against known solutions with cargo test. Vitest covers the interface logic, and Playwright runs whole journeys on desktop and mobile.
Scenes are organised by area of physics and by level. Each has parameters with units, live graphs, the equations behind it and a guided explanation.
Audio to sheet music
Turns audio recordings into editable sheet music, all in the browser: neural pitch detection on TensorFlow.js, key and chord estimation, export to PDF, MusicXML and MIDI. React and TypeScript, nine languages, cross-browser Playwright tests. The score can be edited, played on 35 sampled instruments and practised with falling notes.
Transcription runs on TensorFlow.js in the browser, so recordings are not uploaded. Scores stay on the device and are shared through a link that carries the whole score.
Amadeo estimates the tempo, meter, key and chords, gives each hand its own line and engraves the score with VexFlow.
Amadeo is live in nine languages, with a Pro plan paid through Stripe, a repertoire of arrangements, a tuner and a metronome. An Android app is built and waiting for its store release.
Desktop app for AI coding agents
Desktop app that sends requests to a separate AI coding agent for each project. Data is encrypted at rest, updates are signed, and CI runs on Windows, macOS and Linux. Node.js and Electron.
Each project has its own AI coding agent. The app decides which project a request belongs to and sends it there.
Stored data is encrypted through the operating system’s own keystore.
Updates are signed, and continuous integration runs on Windows, macOS and Linux.
Orion Benefícios, LÍDER Plus Telemedicina and an esthetics studio
Company websites for Orion Benefícios and LÍDER Plus Telemedicina, a website for an esthetics studio in Santos, and an employee portal. The company websites score 95 on mobile and 100 on desktop in Lighthouse and meet WCAG 2.2 AA.
Static pages, built with Astro or with a small Python generator, served by Cloudflare Pages. The company websites score 95 on mobile and 100 on desktop in Lighthouse.
The company websites are tested against WCAG 2.2 AA in all three browser engines: Chromium, Firefox and WebKit.
The employee portal runs on Cloudflare Workers with a D1 database, behind Content Security Policy headers.
Backtesting trading strategies
Backtests trading strategies on out-of-sample data with walk-forward windows, fees and slippage, and uses block-bootstrap tests to check that results are not due to chance. Python, pandas, FastAPI and SQLite, with 221 pytest tests.
Each strategy is fitted on one window and traded on the next, then the window slides forward. Fees and slippage are paid on every trade.
Trying many strategies makes some look good by chance. Block-bootstrap tests check that each result is not due to chance.
Written in Python with pandas, served through FastAPI and stored in SQLite, with 221 pytest tests.
