Calazans Macchiutti

About Me

I started working with data during my undergraduate studies in Physics at UERJ, where I participated in an acoustics research project supported by FAPERJ. My work involved measuring sound levels and studying how material properties affect acoustic reflection. Working with logarithmic measurements such as decibels taught me an early lesson that continues to influence how I approach data: understanding what a measurement represents is as important as the calculation itself.

After graduating, I worked as a physics teacher. That experience helped me develop a skill that later became important in data science: communicating technical concepts clearly to people with different backgrounds and levels of technical knowledge.

I continued my academic training at CBPF, completing a master's degree as a CAPES scholar and a PhD supported by a FAPERJ Nota 10 scholarship. During my doctorate, I also spent a period at the University of Warwick in the United Kingdom as a visiting researcher. I completed my PhD in 2023, receiving an honorable mention for the thesis. My research background gave me a strong foundation in experimental analysis, statistics, signal interpretation, and distinguishing meaningful effects from noise.

I later moved into industry, applying these skills to practical problems. At BINAHKI, I worked on machine learning solutions for industrial clients, including models deployed on AWS using electrical sensor data from equipment such as furnaces and compressors. In one project, the analysis also helped identify a methodological issue in an existing process that was contributing to unnecessary energy consumption.

At ESPM, where I currently work, I have been helping structure data and analytics capabilities within an area that previously relied heavily on spreadsheets and manual processes. My work includes data pipelines, machine learning, business intelligence, and the development of internal generative AI solutions. I also work on evaluation frameworks to assess model quality, cost, and operational viability before deployment.

What I work with

• Machine learning in production: forecasting, anomaly detection, classification, deployment, and monitoring
• Time series and signal analysis: pattern recognition, noisy data, weak signals, and condition monitoring
• Generative AI and LLM systems: RAG, agentic workflows, evaluation pipelines, and quality measurement
• Data products and decision support: KPI frameworks, dashboards, data pipelines, and translating analytical results into practical decisions

I am particularly interested in problems where data, statistical reasoning, and engineering need to work together. I enjoy building solutions that are technically sound, understandable to stakeholders, and useful in real-world decision-making.

Academic Profile

Research in Condensed Matter Physics and Materials Science

Research Focus

During my doctoral program, I focused my studies on the physical properties of transition metal oxides, with a special emphasis on the investigation of double perovskite-type structures and their impact on the "exchange bias" phenomenon.

My research covers a variety of techniques:

• Magnetic measurements
• Electrical and thermal transport studies
• X-ray diffraction
• Magnetoresistance analysis
• Investigation of strong electronic correlation phenomena

Materials Synthesis Experience

Extensive experience in materials synthesis through various methods:

• Sol-gel synthesis
• Co-precipitation methods
• Solid-state synthesis
• Floating zone techniques

All syntheses included both polycrystals and single crystals of ternary and quaternary rare earth oxides, as well as binary single crystals such as V₂O₅.

Characterization Techniques

• X-ray photoelectron spectroscopy (XPS)
• Raman spectroscopy
• Atomic force microscopy (AFM)

Early Research

• Musical Acoustics and Environmental Acoustics
• Applications of electromagnetic waves in magnetic resonance
• Analysis of sound waves in music

View all publications →

Data Science

Professional Summary

My journey has been shaped by the development and application of statistical models, with a strong focus on identifying patterns, understanding trends, and generating actionable insights.

I developed a strong foundation in scientific problem-solving, quantitative modeling, and rigorous analytical reasoning through end-to-end research projects. This experience includes hypothesis formulation, experimental planning, data acquisition, and the interpretation of complex multivariate phenomena. With this background, I present below a detailed overview of my professional trajectory from a technical perspective.

Technical Skills

Programming & Data Manipulation

• Python (Pandas, Polars, NumPy)
• SQL for querying, transformation, and data management
• PySpark and Databricks for large-scale distributed processing

Data Visualization & Statistical Analysis

• Matplotlib, Seaborn, Plotly
• SciPy and Statsmodels for statistical modeling and inference

Machine Learning & AI

• TensorFlow, Keras, Scikit-learn, XGBoost
• Prophet and Darts for forecasting and time series modeling

Cloud & Ops

• Docker for containerization and portability
• AWS, Azure and Google Cloud for scalable cloud architectures
• MLflow for experiment tracking and model registry
• GitHub for version control and collaboration
• Hugging Face for model deployment and sharing

Computer Vision & Media Processing

• OpenCV for image and video processing
• MoviePy and FFmpeg for video automation
• Pillow for image manipulation

Generative AI

• LLMs (Claude, ChatGPT, Gemini) - Advanced proficiency

Professional Experience

Publications

2025

2024

2023

2022

2021

2020

2019

2017

Projects

View all repositories on GitHub

Auditable Agentic RAG

A retrieval-augmented agent built for regulated use: it refuses without evidence, never answers outside a user's scope, and writes every decision to an audit trail

Figures & Results

Two gates and one ledger

18/18, zero scope leaks

Project Overview

Most RAG demos answer well and stop there. In a bank or an asset manager, answering well is table stakes - the questions that decide deployment are different: did it refuse when it had no evidence? could it leak one client's data into another's answer? can you reconstruct, months later, why it said what it said? This pipeline is built around those three.

The Central Design Decision

The model explains; it never computes and never decides. Every figure must appear literally in a retrieved passage, and every claim carries its source. Swap the model and the numbers do not move - only the wording. Scope works the same way: permissions become a metadata filter applied during retrieval, so out-of-scope text never reaches the model. Restricting by prompt is a request, not access control.

Main Products / Deliverables

• Agentic RAG pipeline - retrieve → reason → answer → self-check, structured output, hard budget of two LLM calls per question.
• Guardrails with four refusal decisions - off-policy request, no evidence, no citation, fabricated citation. Off-policy is caught before any LLM call is spent.
• Append-only audit trail - passage IDs, prompt version, model, user, decision. Stores the hash of each passage, never the text.
• Two-layer eval harness - deterministic oracles first, LLM judge only on what was answered.

Results & Impact

Against a keyword floor using the same scope filter, the system scored 18/18 with zero scope leaks. The row that justifies the model is refusal for lack of evidence, 0/3 → 3/3: a search always finds something and answers - refusing requires comprehension, not retrieval. Scope isolation, by contrast, held even at the floor, because the restriction is structural. Judge metrics over the 10 answered items: faithfulness 1.000, answer relevancy 1.000, context precision 0.243 - the last reported rather than hidden, since it measures retrieval noise on a deliberately small corpus.

Full case study View on GitHub ← Back to Projects

Handwritten Document RAG

A compact Brazilian-Portuguese HTR model trained from scratch, evaluated against alternative architectures, and integrated as a pluggable reader in an end-to-end document-RAG pipeline

Figures & Results

Benchmark: 9.57% CER

Architecture

Progression

Project Overview

The core system is an agent that structures PDFs for retrieval-augmented generation. Printed-text OCR handles typed pages, but handwriting it cannot - and no public HTR model exists for Brazilian Portuguese. So I built the handwriting reader from scratch and plugged it in behind the same interface as the existing OCR, selectable at runtime. Nothing downstream needs to know whether a page was typed or handwritten.

Main Products / Deliverables

• pt-BR HTR model - a versioned artifact (weights + alphabet, sha256), 4.3M params, 17 MB; CPU-friendly inference, no GPU required to serve.
• Pluggable reading backend - a drop-in OCR/HTR reader behind available() + ocr_blocks(), swappable with Tesseract without touching the pipeline.
• Document-to-answer pipeline - PDF structuring, vector retrieval, and an agentic RAG loop that turns scanned documents into cited answers.
• Reproducible bakeoff harness - the evaluation and promotion-gate system that picks models by metric, plus a parametrized Colab notebook for managed GPU training.

Key Questions Answered

• Can a competitive Portuguese handwriting reader be trained with no pre-existing model and no local GPU?
• Which architecture actually wins - decided by a metric on the official split, not by intuition?
• How does a research model become a swappable component of a cloud pipeline?

Methodology - how the 9.57% was reached

The number is the output of a disciplined, reproducible procedure. The metric was written before the models: CER is aggregated across the corpus (edits and lengths summed before dividing, as in the ICDAR 2024 evaluation) so the result is directly comparable to published work.

1. Fixed the ground rules. Official BRESSAY splits (19,565 train / 4,609 validation lines), never a custom partition. The alphabet (96 classes) is built from the training split only and travels with the checkpoint.
2. Framed a bakeoff with a promotion gate. A challenger only replaces the baseline if it beats its validation CER. Baseline: CRNN + CTC (no autoregressive decoder, so it cannot hallucinate). Challenger: a fine-tuned TrOCR.
3. Trained with stable, cheap choices. Adam, ReduceLROnPlateau on validation CER, gradient clipping, and the checkpoint saved is the best epoch, not the last; resumes are anchored so a noisy epoch never overwrites a good checkpoint.
4. Read the curve, not the folklore. The epoch 8 - 10 oscillation was not overfitting but a learning rate too high - cutting it 1e-3→5e-4 moved crnn-small's 15.40% down to crnn-large's 9.57%.
5. Evaluated at three levels. Line CER guides training, the segmenter count is checked against ground truth, and page-level CER is measured end-to-end.

How the Technologies Were Used

• PyTorch - the CRNN reader (CNN → bidirectional LSTM) and a custom training loop with best-checkpoint saving, LR scheduling, and anchored resume.
• CTC loss - alignment-free line recognition with greedy decoding; the reason the baseline is trustworthy.
• Hugging Face Transformers / TrOCR - the challenger, fine-tuned with AMP/fp16, LR warmup + cosine decay, and gradient accumulation on GPU.
• PyMuPDF & Tesseract - PDF structuring and the printed-text OCR backend, behind the same contract as the HTR reader.
• Qdrant + FastAPI + Claude - vector retrieval, service layer, and an agentic RAG loop that cites its source.
• pytest & Colab - 105+ tests (hermetic and smoke) and a parametrized notebook for managed GPU training.

Technologies

Getting Started

Results & Impact

A Brazilian-Portuguese handwriting reader that stands among published academic baselines despite CPU-only training (9.57% line CER, beating Flor and Bluche), delivered as a 17 MB versioned artifact ready to ship as a small containerized service. Because every stage is isolated by contract, the recognizer can be swapped - the moment the TrOCR challenger clears the gate - without touching the rest of the pipeline.

Full case study View on GitHub ← Back to Projects

Logistic Data Assessment

Exploratory and predictive analytics for operational efficiency in logistic networks

Screenshots

Departure Delays

Seasonality

Model Insights

Project Overview

This project consolidates exploratory analysis, KPI monitoring, and predictive modeling for a logistic operation handling thousands of daily deliveries. The objective is to surface actionable insights for dispatch planning, identify systemic bottlenecks, and forecast risk factors that affect service-level agreements.

Key Questions Answered

• What routes, hubs, and shifts are the main contributors to departure delays?
• How do seasonal patterns and workload peaks affect fleet punctuality?
• Which operational variables most influence delay risk according to the predictive models?

Analytical Highlights

• Interactive dashboards covering delay distribution, seasonal effects, and operational KPIs
• Machine learning pipeline (gradient boosting and logistic regression) for delay risk scoring
• Explainability layer with feature importance, SHAP interpretation, and scenario simulation
• Automated data quality checks and anomaly alerts for incoming logistic batches

Workflow

1. Data ingestion from transactional systems and IoT telemetry with validation routines
2. Feature engineering combining route metadata, fleet schedules, weather, and customer SLAs
3. Model training and cross-validation with monitoring for drift and performance decay
4. Packaging of results into dashboards and API endpoints for operational teams

Technologies

Results & Impact

The combined analytical assets contributed to a measurable reduction in average departure delays during pilot operations, while the predictive service surfaced upcoming high-risk routes for review. Operational teams now rely on the dashboards to reallocate resources earlier and plan contingencies for peak seasons.

View on GitHub ← Back to Projects

Water Treatment Monitoring

Monitoring system for water treatment plants: forecasting, five anomaly-detection methods, alerting and reports

Screenshots

Main Dashboard

Analytics View

Visualizations

Project Overview

Real-time monitoring for three water treatment plants, each with its own sensor set. Around 7,900 lines across modules and dashboard pages: forecasting with Random Forest, XGBoost, Prophet and ARIMA; anomaly detection by five independent methods; alerting over email, Slack and Teams; automatic PDF, Excel and HTML reports; plus audit and backup. The dashboard runs on synthetic data generated for demonstration; a separate repository holds real SCADA data from the same utility.

Features

• Interactive data visualization dashboards
• Machine learning model demonstrations
• Real-time data analysis tools
• User-friendly web interfaces
• Responsive design for various screen sizes

Applications Included

Technologies

Getting Started

View on GitHub ← Back to Projects

Voltage Time Series LSTM

LSTM neural network for voltage time series prediction and analysis

Screenshots

LSTM Model

Predictions

Training

Project Overview

This project implements a Long Short-Term Memory (LSTM) neural network for voltage time series prediction and analysis. The model is designed to forecast voltage patterns and detect anomalies in electrical systems using advanced deep learning techniques.

Features

• LSTM neural network architecture for time series forecasting
• Voltage pattern recognition and prediction
• Real-time anomaly detection capabilities
• Interactive data visualization and analysis tools
• Model performance evaluation and metrics
• Scalable architecture for different time series lengths

Key Components

Technical Details

Model Architecture

• Multi-layer LSTM with configurable hidden units
• Dropout layers for regularization
• Dense output layer for regression
• Adam optimizer with adaptive learning rate

Data Processing

• Time series windowing for sequence generation
• Min-Max normalization for stable training
• Train/validation/test split with temporal ordering
• Feature scaling and inverse transformation

Technologies

Getting Started

View on GitHub ← Back to Projects

Fraud Detection (KYC/KYT)

Data Science approaches for detecting fraudulent transactions using supervised and unsupervised learning

Screenshots

📄 Research Paper (click to download)

PCA Features

Model Comparison

Amount Analysis

Project Overview

This project implements a comprehensive fraud detection system combining Know Your Customer (KYC) and Know Your Transaction (KYT) methodologies. Using the Credit Card Fraud Detection dataset, the system evaluates multiple machine learning approaches including supervised classifiers, unsupervised anomaly detection, and deep learning autoencoders to identify fraudulent transactions in highly imbalanced datasets.

Key Questions Answered

• How do supervised models (XGBoost, Random Forest) compare against unsupervised approaches?
• Which PCA-transformed features show the strongest discriminative power?
• What is the optimal threshold strategy for extreme class imbalance (~0.17% fraud rate)?
• How can autoencoder reconstruction error be leveraged for anomaly detection?

Analytical Highlights

• XGBoost achieved highest Average Precision (0.858) followed by Random Forest (0.826)
• ROC-AUC scores above 0.96 for all models, with XGBoost reaching 0.969
• Autoencoder-based anomaly detection with tuned threshold achieving AP of 0.510
• Feature importance analysis revealing V14, V4, V12, V10 as top discriminators

Technologies

Results & Impact

Supervised learning significantly outperforms unsupervised methods when labeled data is available. XGBoost emerged as the best performer with 85.8% Average Precision, detecting 86% of fraudulent transactions while maintaining 51% precision. The autoencoder approach provides a viable alternative for cold-start scenarios when labeled fraud examples are scarce.

Download Paper View on GitHub ← Back to Projects

Curriculum Vitae

Experience

Education

Skills

Links

GitHub LinkedIn Lattes Scholar