Gustavo Barra
Gustavo Barra Felizardo
Computer Science student at UFMG. Undergraduate researcher working on LLM optimization and SLM applications. Building Solitus and Junto Finance on the side.
I don't just study AI. I deploy it.
About me
Gustavo Barra
CS UFMG · Researcher & Founder
I'm Gustavo Barra Felizardo — a Computer Science student at UFMG and Undergraduate Researcher at FutureLab DCC/UFMG in partnership with Morada AI, elected Brazil's #1 AI startup by Top 100 Startups. I work with SLM and LLM optimization techniques such as QLoRA, PEFT, Knowledge Distillation, and Fine-Tuning.
I'm also the Co-Founder of Solitus, a B2B2C events ecosystem, and the founder of Junto, an AI-powered iOS finance app with on-device ML. I'm a national olympiad medalist (OBMEP, OBA, ONC) and did a high school exchange semester in Vancouver, Canada.
0+
Projects
0
Building Startups
0
Finalizing Papers
Journey
2023
Exchange Semester in Vancouver
High school exchange semester at Burnaby Mountain Secondary School in Vancouver, Canada. Sharpened English, broadened perspective, and came back ready to build.
2025
UFMG + FutureLab Research
Joined Computer Science at UFMG and one month later became Undergraduate Researcher at FutureLab, working on LLM optimization techniques and Lead classification for Morada AI — Brazil's #1 AI startup.
Now
Solitus, Junto & Research
Co-founding Solitus (B2B2C events platform), building Junto (AI iOS finance app with on-device ML), and co-authoring 2 research papers on LLM optimization and Lead Classification pipeline.
Vision
I believe the world is shaped by founders and engineers willing to take risks, which is why I like to get involved with ambitious projects that create deep impact in a niche.
Human neuroplasticity decreases over time, the window to build deep skills is shorter than it feels. That keeps me always in motion: exchange in Canada, one of Brazil's most demanding CS programs at UFMG, two businesses in parallel. Each one forced me to think in ways I couldn't before.
Build things that matter.
Projects
Here's what I've been building.
Co-Founding · Events Ecosystem Platform
B2B2C platform for the live events industry. Unifies ticketing, cashless payments, parking management, stock control, consuming-software and behavioral analytics into a single ecosystem spanning a producer app, ticket validation, and consumer-facing app.
Founder & Full-Stack Engineer · iOS Finance App
iOS app combining on-device ML and cloud services for intelligent expense tracking and financial insights. Uses MLX-Swift with QLoRA-finetuned models for local inference and integrates Open Finance APIs for automated bank syncing. Preparing for App Store beta release.
End-to-end SLM implementation based on Dr. Raj Dandekar's (MIT PhD, ML) teachings. Covers the full pipeline: dataset creation, tokenization, input–target pair construction, model architecture design, pre-training setup and execution, and inference. Also explores the theory behind it: Backpropagation, activation functions (ReLU & GeLU), and the step-by-step reasoning behind each architectural decision.
Social network built from scratch in C++11 with SQLite3, using OOP principles and MVC architecture. Features real-time score ranking, bidirectional follow relationships, and gamification mechanics with robust exception handling.
Space Invaders game built from scratch in C, using ALLEGRO library . Features classic gameplay mechanics, enemy movement patterns, and power-up system.
XGBoost model for real estate lead scoring using Target Encoding, Feature Engineering, and 5-fold Cross Validation. Achieved 53% RMSE improvement over baseline by engineering features like purchase_power and affordability ratio, eliminating profession-based overfitting.
Neural network built from scratch with NumPy for real estate lead profile scoring. Implements forward and backward propagation, gradient descent, and activation functions without any ML framework — covering the full pipeline from raw data to prediction.
Bibliometric analysis of 12,209 scientific publications on LLM and Knowledge Distillation, sourced from Scopus. Built co-occurrence networks in VOSviewer and plotted publication trends in Python — mapping where the research field is heading and what emerging terms are gaining traction.
Posts & Ideas
Perspectives from research, building business, and learning things the hard way — from fine-tuning LLMs to shipping iOS apps.
Mar 06, 2026
From an Nvidia article to hands-on QLoRA fine-tuning at UFMG's FutureLab — why I believe SLMs are the most important shift happening in AI right now.
Mar 06, 2026
How I built a privacy-first finance app using local LLM inference with MLX-Swift, QLoRA fine-tuning, RAG, and Open Finance APIs — and the real tradeoffs nobody talks about.
Contact
Open to discussing research collaborations, startup ideas, technical questions, or anything about AI, product and technology.