João Assaoka

• Regular Direct PhD Student (Jan 2027 – Dec 2030): Invited to enter directly into the PhD program in recognition of academic performance and scientific production during undergraduate studies.
• Special Graduate Student (Mar 2024 – Present): Completed graduate courses concurrently with undergraduate studies, earning the maximum grade (A) in all: Algorithm Analysis and Data Structures, Artificial Intelligence, Neural Networks, Operations Research, and Seminars.

• Member researcher of the DNLP Group and Malanis Lab, participating in discussions and experiments on the state of the art in Natural Language Processing (NLP) and Machine Learning.

• Ranked 3rd place in the career progression selection process following the interdisciplinary cycle.
• Relevant Coursework: Calculus, Linear Algebra, Linear Optimization, Operations Research, Problem Solving via Mathematical Modeling, Topics in Economics and Business, Algorithms and Data Structures, Object-Oriented Programming, Software Engineering, Databases, Probability and Statistics, Inference and Regression Analysis, Machine Learning, Artificial Intelligence, and Neural Networks.

• Certificate of Emphasis in Data Science Fundamentals.
• Academic Teaching Assistant: Served as TA for Programming Logic and Algorithms & Data Structures courses.

• Increased correlation with human judgment by 113% (from 0.23 to 0.49) in reference-free financial news summarization evaluation, proposing and implementing task decomposition via Question Answering with LLMs in Python and LangChain.
• Accelerated multi-task retraining cycle by 99.5% (from 1h17min to 25s) while preserving 94.7% of Macro-F1, by architecting an SLM Model Merging strategy via Mergekit.
• Increased relative Macro-F1 by 31.2% in multi-class toxicity detection under imbalanced scenarios by structuring synthetic data augmentation pipelines via LLMs.
• Built a proprietary financial domain dataset (5,000+ samples), developing scraping with Selenium, a web annotation platform in Streamlit, relational modeling in PostgreSQL (Neon/SQLAlchemy), and exploratory analysis with Pandas and Matplotlib.
• Architected fine-tuning pipelines (Full, LoRA, IA3, Soft Prompts via Hugging Face) and inference (vLLM) with structured outputs, orchestrating experiments across multi-GPU NVIDIA nodes in an HPC Linux environment (Digital Research Alliance of Canada) via Slurm.

• Founded and structured the consultancy's dedicated Data Science and Artificial Intelligence department, leading the spin-off from the computing division and establishing new hiring criteria that enabled talent attraction across multiple engineering fields.
• Managed the lifecycle and technical delivery of 6 projects, encompassing 3 consulting initiatives for external clients and 3 internal operational automation products focused on process scalability.
• Upskilled new consultants in Data and AI competencies, designing the technical learning track and conducting training sessions in Python, exploratory data analysis (Pandas, Matplotlib), CRISP-DM methodology, interactive Streamlit interfaces, and LLM orchestration with LangChain.

• Reduced clinical report generation time by over 90% (from 1–3 hours to 5 minutes) by architecting an integrated data solution with a Streamlit interface and an automated LLM summarization pipeline.
• Developed an intelligent conversational agent (Text-to-SQL / Chatbot) capable of interpreting natural language queries and extracting on-demand analytical metrics from cognitive workshop data and elderly participant medical records.
• Centralized and standardized the workshop data ecosystem, replacing fragmented legacy spreadsheets with a structured relational architecture and input forms.

• Enabled operational cost savings and proactive industrial failure prevention by predicting sensor time-series degradation signals, developing and benchmarking models ranging from linear baselines (Linear Regression) to gradient-boosted ensembles (XGBoost) and deep recurrent architectures (LSTM).
• Implemented a multi-technique Explainable Artificial Intelligence (XAI) framework using SHAP, ALE (Accumulated Local Effects) plots, and partial dependence (PDP 1D/2D), diagnosing complex causal interactions to validate prediction alignment with domain engineering knowledge and increase decision-making transparency.
• Co-created the IAsmin educational platform, democratizing practical XAI and Industry 4.0 predictive maintenance methodologies in Portuguese through 7 hands-on tutorials detailing theoretical foundations and interpretability implementations.

Reduced multi-task model creation time to just 0.5% of standard fine-tuning by merging specialist LLM checkpoints without performance degradation. Joint work during research visit with Dalhousie University.

Study analyzing reference-free summarization quality metrics for financial news in Portuguese, demonstrating that question-answering decomposition outperforms monolithic LLM judges against human annotators.

Investigation into data scarcity and class imbalance in toxic speech detection, evaluating LLM-based synthetic generation and cross-dataset label propagation strategies in Portuguese.

Multimodal architecture combining textual and visual signals for detecting financial misinformation and deceptive news in Brazilian Portuguese.

Extensive benchmark of neural architectures evaluating both in-domain performance and out-of-distribution robustness across economic news corpora.

Multi-LLM negotiation framework to improve classification precision and generate faithful rationales for toxicity detection in Portuguese.

• Winner in the Technical or Technological Products category for outstanding undergraduate research and scientific development.

• Project selected in the campus stage under the Exact Sciences Scientific Research category to represent the campus in the university-wide institutional award (pending).