Leandro Seixas

Leandro Seixas

he/him

Physicist · Professor · Full Stack Developer · AI Engineer

Portfolio

Software built around a simple idea: the hard part of computational materials science should be the science. The research code lives in seixas-research, applied data science and ML engineering in seixas-solutions, teaching material in seixas-teaching, and everything built for the joy of it in seixas-fun.

Poraquê

Machine learning density-based operators.

Poraquê is a software framework for machine learning operators between the real-space fields of density functional theory. Given only a crystal geometry, it predicts the charge density and the kinetic energy density. No wavefunctions, no self-consistency cycle.

  • Two Fourier Neural Operators: the Hohenberg–Kohn map, and the kinetic energy density functional missing from orbital-free DFT
  • Density-based operators as first-class objects
  • Built for the standard scientific Python stack (PyTorch, ASE)
PythonPyTorchIn development
WebsiteManualPyPISource

Calango Studio

Visual atomistic modeling.

A desktop application that brings three things into one window: an interactive viewer for building, editing and inspecting atomic structures; a calculator-agnostic environment for setting up and running simulations; and viewers that turn the output back into figures you can read.

  • Builders for slabs, interfaces, dislocations, polycrystals, nanotubes, nanoribbons and special quasirandom structures
  • Materials Project, PubChem and C2DB one search away
  • Written in C++20 with Qt 6, responsive at tens of thousands of atoms
C++Qt 6In development
WebsiteManualSource

Mandacaru

Fermionic quantum simulation, one API.

A lightweight Python framework for fermionic quantum simulation based on variational quantum algorithms. From a molecular or periodic geometry, it builds real-space grids, evaluates the one- and two-body integrals, maps the Hamiltonian to qubits, and solves it variationally — through a single ASE calculator that runs unchanged on IBM Qiskit, Amazon Braket and Google Cirq.

  • VQE, ADAPT-VQE and excited states — deflation and subspace search — sharing one driver
  • Localized basis sets and pseudopotentials generated from scratch, not tabulated
  • Cross-backend agreement to 1.3×10⁻⁷ Ha, validated on real quantum hardware
PythonQuantum computingIn development
WebsiteManualPyPISource

Research tools

Smaller packages that come out of the research directly, published on PyPI under seixas-research.

Onça-pintada

Thermodynamics of alloys.

Computes enthalpy, entropy and Gibbs free energy of mixing for binary and multicomponent alloys within the subregular solution model, together with binodal and spinodal curves. Cluster interactions are evaluated in the quasi-chemical approximation, which lets it describe chemical short-range order and rebuild atomistic geometries from short-range order metrics.
PythonAlloysThermodynamics
ManualPyPISource

SAGUI

Machine-learned interatomic potentials and generative models.

Scalable Atomistic Graph networks for Universal Interactions: an equivariant graph neural network framework that learns energies and forces from the same structures it learns the distribution of structures from. Two interchangeable architectures — a many-body ACE-style message-passing model and a strictly local one — with the O(3) tensor algebra implemented from scratch, no e3nn dependency.
PythonPyTorchMLIPIn development
Source

Blendpy

Alloy thermodynamics from first principles.

A toolkit for investigating thermodynamic models of alloys with first-principles calculations: enthalpy of mixing and the spinodal and binodal decomposition curves of a phase diagram, through the dilute solution interpolation model, driven by any ASE calculator.
PythonASEPhase diagrams
PyPISource

Quasigraph

Descriptors for materials machine learning.

A graph-like chemical and geometric descriptor toolkit: from an ASE Atoms object it produces a dataframe or a vector combining a chemical part and a coordination-number-based geometric part, ready to feed a machine learning model.
PythonDescriptorsMaterials informatics
DOIPyPISource

calango-cli

Calango workflows, headless.

Runs a pipeline built on the Calango orchestration canvas without a display: export the workflow to a single self-contained JSON document, copy it to an HPC cluster, run it there, and load the results straight back into the GUI.
PythonHPC
Source

Data science & ML engineering

Applied work under seixas-solutions.

Sucuri

Anomaly signals in Brazilian federal spending on higher education.

Collects, cleans and analyses budget execution from the Ministry of Education through the Portal da Transparência API — federal universities, institutes, university hospitals, CAPES, FNDE/FIES — enriched with contracts, procurement, sanctions and IBGE population and GDP data. It flags statistically atypical spending patterns worth a manual check; statistical atypicality is not evidence of irregularity.
PythonOpen dataAnomaly detection
Source

Papagaio

Local-first AI speech refinement for Apple Silicon.

Transcribes spoken audio with word-level timestamps, detects filler words, stutters, repetitions, dead air and rushed delivery, and renders a cleaned-up version that still sounds like the same person in the same room. Everything runs on the machine, with MLX — no uploads, no API keys.
PythonMLXWhisperAlpha
Source

For fun

Tamanduá

Can you guess all the flags?

A flag-guessing game across national flags, 2026 World Cup teams and the 26 Brazilian states plus the Federal District, in multiple-choice or type-the-name mode.
PythonWeb app
PlaySource

Teaching material

Notebooks and tutorials used in courses are published in seixas-teaching and listed on the CV.

Have a problem worth computing? I am open to research collaborations and consulting on scientific software and applied machine learning — start a conversation.