About meMy ProjectsDico the agentMy MusicReading list
My Projects
Address
My Projects
7 object(s)
My Music

The GOATs

100 tracks · 7:48:09 · by Felipersas

Spotify

Double-click a track to play it.

Doom

DOOM

The first episode of DOOM (1993), in the shareware release id Software let people distribute, running in DOSBox inside the browser.

Arrows: move · Ctrl: fire · Space: open doors · Shift: run · Esc: menu

DOOM © id Software, unmodified shareware v1.9. Emulated by js-dos (GPL-2.0). The game and the emulator are a few MB and load only when you click.

Felipe Marques – Browser
Felipe Marques

Felipe Marques

Software Engineer

Mid-Level Full-Stack Developer @ Keeper · Sorocaba, SP, Brazil

About

I'm a software engineer focused on applied AI, with a full-stack base: backend, frontend, mobile and the infrastructure that runs it.

At FazzLeads, a company I co-founded for real estate agencies, I put into production an AI agent that answers and qualifies leads on WhatsApp and now books around 200 property visits a month. It runs on Mastra AI and MCP, behind guardrails, metrics and alerts. This portfolio has an agent of its own, grounded in the site's data and measured with evals. Before that, at MindGroup, I architected and shipped 3 production systems and took a critical query from 5s to 240ms. I'm at Keeper today, a fintech for graduation fundraising.

I'm studying Machine Learning (DeepLearning.AI certificate and a Systems Analysis degree at SENAI). Outside of work I build distributed systems and developer tools, including DemoCraft.

Experience

Keeper

Location
São Paulo, SP, Brazil
Location type
(Hybrid)

MindGroup Consulting e Marketing

Location
Sorocaba, SP, Brazil
Location type
(Hybrid)

FazzLeads

Location
Remote

Projects(7)

Portfolio Agent

A durable, bilingual AI agent embedded in this portfolio to answer questions about my career, projects, and technical decisions.

Built with Vercel Eve and OpenAI's GPT-6 Luna, with generative UI rendered through json-render and a local multilingual Whisper pipeline running through Transformers.js in a Web Worker.

  • Eve
  • OpenAI
  • json-render
  • Transformers.js
  • Next.js

DemoCraft

An open-source toolkit for creating polished and reproducible product demos from real application workflows.

Demo flows are defined in TypeScript, executed in a real browser with Playwright, and rendered deterministically with Remotion.

  • TypeScript
  • Playwright
  • Remotion
  • React

Real-Time Crash Game

A real-time multiplayer crash game where every cent a bet moves is posted to a double-entry ledger and every crash point can be verified by the player.

Games and Wallets, two NestJS services, settle bets through a choreographed Saga over RabbitMQ with a transactional outbox and an idempotent inbox. The ledger is SHA-256 chained per wallet and reconciled every 5 minutes.

Tamper-evident double-entry ledger and provably fair crash points

  • NestJS
  • RabbitMQ
  • PostgreSQL
  • Redis
  • WebSockets

PayFlow

A financial transfer system built with three Go microservices that communicate exclusively through asynchronous messaging.

Uses a choreographed Saga with no central coordinator, DDD boundaries, idempotent operations, and OpenTelemetry tracing to keep transfers consistent across distributed services.

  • Go
  • PostgreSQL
  • RabbitMQ
  • Docker

Fraud Detection

A neural network that flags fraudulent credit card transactions, built on 1.85 million simulated purchases and served by a FastAPI service as a versioned 6 KB model bundle.

Trained in Keras with a time-based split, PR-AUC as the metric and a cost-based decision threshold, then exported to ONNX and served with ONNX Runtime, so the API image needs neither TensorFlow nor scikit-learn. Each card's recent purchases live in Redis, shared by every replica.

PR-AUC of 0.959 on the untouched test set, catching 95.8% of frauds

  • Python
  • TensorFlow
  • ONNX
  • FastAPI
  • Redis
  • Docker

Mini-GPT

A decoder-only GPT written from scratch in PyTorch, one component at a time, each with its own tests, an experiment and a document that explains the math.

No nn.Transformer: a BPE tokenizer, causal multi-head attention, pre-norm Transformer blocks and a weight-tied LM head add up to 7,391,744 parameters, trained on CPU, CUDA or Apple's MPS over four Machado de Assis novels and a slice of the Portuguese Wikipedia.

Validation perplexity of 12.82 after a 15-epoch run

  • Python
  • PyTorch

js-routines

Go-style concurrency for TypeScript: channels, select, synchronization primitives, cancellation and a deterministic scheduler for testing concurrent code.

Three modules: a scheduler that runs routines written as generators in an order set by a seed, for tests; async/await on the event loop, for production, with the same channels and primitives; and channels and a mutex on SharedArrayBuffer, for parallelism across workers.

A seed reproduces every time a race that shows up in 4.6% of runs

  • TypeScript
  • Bun
  • Node.js

Stack

Education

SENAI Sorocaba

Period
Jan 2025 — Present

Associate Degree, Systems Analysis and Development

Data Science, Data Analysis, Java, Algorithms, Machine Learning, Software Engineering, Networks

SENAI Sorocaba

Period
Jan 2023 — Dec 2024

Technical Degree, Systems Development

MySQL Database, Networks, TypeScript, Java, OOP, Algorithms

Certifications(2)

Amazon Web Services (AWS)

Jun 2026

AWS Academy Graduate: Cloud Developing (Training Badge)

Verify on Credly

AWS, AWS Cloud Computing, Developing on AWS

DeepLearning.AI

Jan 2026

Supervised Machine Learning: Regression and Classification

ID / L2MYYHXMHCPD

Linear, Multiple & Polynomial Regression, Gradient Descent, Regularization (L1/L2), Logistic Classification, Neural Networks, Practical ML Development

DoneInternet