Igor L.R. Azevedo

RELEVANT EXPERIENCES

1. Research Scholar at The University of Tokyo — Graph Neural Network Laboratory Tokyo, Japan | Apr 2023 – Apr 2025
   • Served as a Research Scholar (a role similar to a Research Assistant) for two years in Prof. Toyotaro Suzumura's lab, contributing to machine learning research on recommender systems, foundational LLMs, and financial forecasting.
   • Led a research partnership with Nikkei Inc., Japan's largest financial media company and owner of the Financial Times, analyzing their data to design and propose novel recommender system models for user behavior (under Dr. Yuichiro Yasui).
   • Built LLM systems including multi-agent pipelines, prompt engineering, and long-context handling with Retrieval-Augmented Generation (RAG).
   • Trained large-scale distributed models on the Wisteria supercomputer.
   • Engineered deep learning models for high-frequency stock price forecasting, with a focus on predicting market behavior during election periods.
   • Implemented Point-of-Interest Graph Neural Network recommendation models during an industry partnership with Toyota.
2. Developer & Development Team Lead at VOGA Brasília, Brazil | Jul 2021 – Apr 2023
   • Started as a Developer, later promoted to Development Team Lead.
   • Led system integration for VOGA post-acquisition by BTG Pactual, Latin America's largest bank.
   • Delivered interactive data solutions and analytics for non-technical stakeholders throughout the acquisition, using FastAPI and Plotly to build self-service dashboards integrated with BTG's API.
   • Architected a centralized, real-time data ecosystem managing over USD 300 million in portfolios.
   • Transitioned from full-stack development (Flask/NextJS) to a DevOps role, managing AWS infrastructure (EC2, VPC, RDS, ELB) with Terraform, automating CI/CD with Jenkins, and securing the application with Cloudflare.
3. Research Intern at Cellcrypt: Quantum-Safe Encrypted Calls London, UK (Remote) | Sep 2020 – Jun 2021
   • Advised by Professor Edson Mintsu Hung and Dr. Rogério Richa, improving VoIP call quality by 7% in voice-encrypted systems through automated ML parameter tuning.
   • Analyzed voice quality metrics (MOS, jitter, latency) under varying network and encryption conditions.
   • Built an ML pipeline for automated PJSIP (multimedia communication library written in C with high level API in C++) parameter tuning, which boosted call quality metrics by 7%.

RESEARCH PAPERS

1. A Look Into News Avoidance Through AWRS: An Avoidance-Aware Recommender System
   January - July 2024 | SDM'25 Proceedings | ArXiv
   • Collaboration: Toyotaro Suzumura (The University of Tokyo) and Yuichiro Yasui (Nikkei Inc.)
   • Highlights: Developed AWRS, an Avoidance-Aware Recommender System for news that incorporates article avoidance as a key factor to improve recommendations. Evaluated on datasets in English, Norwegian, and Japanese, AWRS outperformed existing methods by leveraging avoidance as an indicator of user preferences.
2. NewsReX: A More Efficient Approach to News Recommendation with Keras 3 and JAX
   January - August 2025 | ArXiv
   • Collaboration: Toyotaro Suzumura (The University of Tokyo) and Yuichiro Yasui (Nikkei Inc.)
   • Highlights: A modular and extensible framework for news recommendation systems research, implementing state-of-the-art models with a focus on reproducibility and ease of use. The framework has been optimized with Keras 3 + JAX backend for enhanced performance through JIT compilation and XLA acceleration.
3. POPK: Mitigating Popularity Bias via a Temporal-Counterfactual
   April - July 2024 | ArXiv
   • Collaboration: Toyotaro Suzumura (The University of Tokyo) and Yuichiro Yasui (Nikkei Inc.)
   • Highlights: Developed POPK, a model which uses temporal-counterfactual analysis to reduce popularity bias in news recommendations. POPK improves accuracy and diversity by systematically removing the influence of popular articles.
4. From Votes to Volatility Predicting the Stock Market on Election Day
   August - December 2024 | ArXiv
   • Collaboration: Toyotaro Suzumura (The University of Tokyo)
   • Highlights: Deep learning models for high-frequency stock price forecasting, focused on predicting market behaviour during election periods.
5. A SHA-3 Co-Processor for IoT Applications
   January - November 2020 | Paper (IEEE - WCNSPS'20)
   • Collaboration: Alexandre S. Nery (University of Brasília) and Alexandre da C. Sena (Rio de Janeiro State University)
   • Highlights: Designed and implemented a SHA-3 hardware co-processor on FPGA for IoT applications, achieving 65% faster performance than ARM Cortex-A9 with improved energy efficiency and reduced circuit area.

AWARDS

1. 3rd Place – HackTheLaw, King's College Cambridge June 2026 | About
   About: Hosted by the University of Cambridge and Stanford CodeX. Recognized for technical excellence and LLM integration by a panel drawn from academia, industry (Perplexity, Anthropic), and law firms (Clifford Chance and White & Case).
2. SIAM Travel Award – SDM25 May 2025 | About
   About: Granted by the Society for Industrial and Applied Mathematics (SIAM), this award supported travel to the 2025 SIAM International Conference on Data Mining (SDM25) held in Alexandria, VA. It recognized promising early-career researchers contributing to the field.
3. Japanese Government (MEXT) Research Scholarship April 2023 - April 2025 | About
   About: The Japanese Government (MEXT) Research Scholarship supports international students conducting research at Japanese higher education institutions. Approximately 9,000 students are accepted out of 100,000 global applicants (~9% acceptance rate), with full funding for tuition, living expenses, and research activities.
4. Brazilian Government (CNPq) Institutional Scientific Initiation Scholarship (PIBIC) August 2019 - July 2020 | About
   
   About: The PIBIC program, funded by the Federal Government of Brazil, supports undergraduate students in research, technological development, and innovation. Awarded to the top ~4% of undergraduates nationwide on academic merit; funded research into FPGA-based cryptographic security systems.

CERTIFICATIONS

• IELTS: Overall Band Score: 8.0 (Listening: 8.0, Reading: 8.5, Writing: 7.0, Speaking: 8.5)

OPEN SOURCE CONTRIBUTIONS

1. NewsReX - Code | Weights
   An open-source framework for neural news recommendation, built with The University of Tokyo and Nikkei. I am the first author and wrote the entire implementation, and have maintained it for over a year.
   
   It is research tooling: models are declared in YAML and composed through Hydra, and run on two backends — PyTorch and JAX/Flax — behind a single interface. Cross-framework parity tests verify that both implementations agree numerically before any result is reported, and evaluation is shared and framework-agnostic. Everything is MIT-licensed and every trained weight is public, so results are reproducible by inference alone.
   
   Shared by François Chollet, the creator of Keras.
2. AlphaDL - Code
   Lightning AI Project for Cutting-Edge Deep Learning Models in Stock Markets.
3. NewsrecLib - Code
   Implemented the PP-REC SOTA model into the news recommendation framework.
4. Qlib - Code
   Added support for the Brazilian stock market, enabling local investors and researchers to use Qlib's machine learning models and data processing pipelines on Brazilian stock data.

ESSENTIALS

COURSES

• Deep Learning @ Carnegie Mellon University | Certificate
• Digital Signal Processing @ EPFL | Certificate
• Information Theory @ The Chinese University of Hong Kong | Certificate
• Algorithms and Data Structures @ UCSD | Certificate
• Model United Nations @ Harvard University | Certificate
• Model United Nations @ Yale University | Certificate

LEADERSHIP EXPERIENCES

• Commercial Director, Electrical Engineering Junior Enterprise (ENETEC) June 2019 - April 2020
  
  Led the team to achieve annual goals set by the National Association of Junior Enterprises in Brazil (Brasil júnior), securing projects worth approximately USD 30,000, which funded training and development for company members.

LANGUAGES

TECH STACK