Matheus Ferraroni Sanches

• Leadership and R&D · Technology strategy, R&D portfolio management and innovation roadmaps, project and resource prioritization, investment planning, building and leading multidisciplinary teams, hiring, managing through team leaders, mentoring and digital transformation.
• Artificial intelligence · LLMs and transformers, fine-tuning, continued pretraining, dataset construction and curation, model evaluation, RAG, AI agents, embeddings, federated learning, differential privacy and data augmentation.
• Applied AI and adoption · Enterprise AI adoption and governance, generative AI for engineering and software, prompt engineering and language model integration with enterprise systems via MCP.
• Multidisciplinary R&D · Product development integrating software, hardware, embedded systems and mechanical engineering, including requirements definition, prototyping and validation.

• Cloud and architecture · AWS, large-scale distributed systems, serverless and event-driven architectures, parallel and concurrent processing, Docker, CI/CD, Linux, messaging, APIs and cloud cost optimization.
• Data and persistence · PostgreSQL, MongoDB, DynamoDB, Snowflake, Redis
• Geospatial processing and agritech · Remote sensing, Landsat and Sentinel-2, large-scale raster processing, GDAL/Rasterio, vegetation indices, spatial statistics, precision agriculture, variable-rate application maps and integration of field data with satellite imagery.
• Programming · Python and JavaScript/TypeScript
• Languages · Native Portuguese · Fluent English

Quero-Quero

Brazilian Portuguese data and continual pretraining

A reproducible research pipeline for preparing six Brazilian Portuguese corpora and adapting Tucano 2 0.6B through continual pretraining.

Problem:

Prepare heterogeneous data at scale and train language models with provenance, privacy, and comparability across experiments.

Contribution:

Designed and implemented validation, conservative cleaning, deduplication, tokenization, packing, verifiable shards, and resumable training with loss and perplexity evaluation.

Current status:

Preparation for all six datasets and the training pipeline are implemented; the complete scientific run depends on local derivatives and cluster allocation.

Technologies: Python, PyTorch, Hugging Face Transformers, Parquet, Slurm, DDP, NCCL.

WackyWacky

Experimental crawler for controlled corpus construction

An open-source web crawler for controlled exploration, compressed text collection, and full execution observability.

Problem:

Crawl pages recursively while retaining control over domains, language, access rate, failures, and stored volume.

Contribution:

Implemented concurrent workers, domain and language filters, rate and retry limits, compressed persistence, and telemetry for metrics, logs, and traces.

Current status:

Functional experimental crawler published as open source.

Technologies: Python, Playwright, MySQL, Docker, OpenTelemetry, Grafana, Loki, Prometheus, Jaeger.

WackyWacky Analysis

External-memory corpus analysis

A standalone package for characterizing immutable WackyWacky corpus snapshots without loading tens of gigabytes into memory.

Problem:

Validate, deduplicate, and characterize a large text corpus with resumable and auditable results on constrained hardware.

Contribution:

Developed inventory, exact deduplication, boilerplate review, lexical and structural analysis, checkpoints, and deterministic table and figure generation.

Current status:

The pipeline and primary B_clean_v2 view are implemented; the complete run executes on the dedicated processing server.

Technologies: Python, Parquet, Zstandard, spaCy, Slurm.

Semantic Replace with Federated Learning

Privacy and personal-data reproduction in federated learning

Experimental research on targeted reproduction of synthetic personal profiles and leakage mitigation in federated language-model training.

Problem:

Measure whether an adversarial client increases sensitive-data reproduction and compare defenses under controlled federated conditions.

Contribution:

Implemented synthetic data, local training, FedAvg, extraction auditing, rotating semantic substitution, per-conversation DP-AdamW, and resumable pilots.

Current status:

The semantic-substitution pilot completed and passed on two seeds; the refined differentially private scientific run remains pending.

Technologies: Python, PyTorch, Hugging Face Transformers, Opacus, Slurm.

Visual Algo

Interactive algorithm demonstrations

A published collection of visual experiments for exploring algorithms, machine learning, and neural evolution directly in the browser.

Problem:

Make abstract concepts observable and adjustable through educational experiences that require no installation.

Contribution:

Created and maintains demos for K-Means, perceptrons, evolved neural vehicles, and deterministic evolutionary simulation.

Current status:

Published on GitHub Pages with live demonstrations.

Technologies: JavaScript, HTML, CSS, p5.js, TensorFlow.js, Chart.js.

RP-Sim

Radio propagation simulator

An open-source environment for simulating wireless communication propagation, with support for propagation models, RSSI, BER, communication-area export, and an interactive demo.

Technologies: JavaScript, p5.js, Mappa.js.

MCCD

Conversational methodology and datasets

A methodology for generating natural conversational datasets from online forums and other sources of human dialogue.

Miner-XenForo

Forum-mining tool

A tool for extracting and preparing conversational data from forums for use in natural language processing datasets.

Technologies: Python.

Portuguese NLP dataset

Datasets for natural language processing in Portuguese

A collection of new or adapted datasets for training Portuguese-language models.

GGA

Python library for genetic algorithms

An open-source library that simplifies the creation and customization of genetic algorithms in Python.

Msim

Event simulator

A simple event simulator written in Python.

Technologies: Python.

Satellite Image Processing

Satellite image processing

Processing Landsat 8 and Sentinel-2 satellite imagery to support farmers’ decision-making with color imagery and NDVI.

Technologies: AWS, Python, GDAL, Rasterio, Pillow.

SCO3, Servidor Clickideia Offline 3.0

FAPESP project

A project focused on automatically updating educational content through opportunistic networks in schools with limited or no internet access.