# Theylor Machado

Data Scientist · Demand Forecasting and Time Series · Claude Certified Developer

At nPLAN, I build demand-forecasting solutions that support planning decisions, connecting models, data and product development. In October 2026 I earned the Claude Certified Developer credential from Anthropic. It certifies developers who build, integrate and ship production applications and agents on Claude, using the Claude API, Claude Code, custom tools and MCP servers.

- Website: https://theylor.dev
- Email: contato@theylor.dev
- LinkedIn: https://www.linkedin.com/in/theylor921
- GitHub: https://github.com/theylor999
- Location: Sapucaia do Sul, RS, Brazil
- CV: [English PDF](https://theylor.dev/documents/cv_theylor_en.pdf) · [Portuguese PDF](https://theylor.dev/documents/cv_theylor_ptbr.pdf)

## Experience

### Data Scientist, NEO Digital Industries · nPLAN

Jul 2025 to present · Porto Alegre, Brazil · Hybrid

Impact:

- Forecast accuracy up at least 10 percentage points: I built automatic per-series parameter and model selection. Multiple multinational companies use it, and it beat the models they were already using by at least 10 percentage points of accuracy. Estimated business impact for those client companies, not measured: about +2% profitability and −5% total inventory cost, derived from IBF (Institute of Business Forecasting) and McKinsey & Company benchmarks (IBF: +0.2% profitability per 1 pp of accuracy; McKinsey: −0.5% inventory cost per 1 pp, or about −5% for 10–20% improvements).
- About 75% faster processing: forecasting runs went from roughly 8 hours to 2 with GPU/CUDA execution, parallelism, batching and caching.

Key contributions:

- End-to-end delivery across Python, C#/.NET, EF Core, PostgreSQL and React/TypeScript, including APIs, domain services, indicators and analytical interfaces.
- Developed more than 20 demand-forecasting models spanning statistical models, Machine Learning, Deep Learning, foundation models and ensembles.
- Holdout, expanding-window and sliding-window backtesting, per-series best-model selection and leakage-prevention controls.
- Evaluation with multiple metrics, including bias and uncertainty, across different demand profiles.
- Experience with multiple exogenous datasets, plus calendar and hierarchy features, with time-aware signal availability.
- Pipeline optimization through batching, caching, parallelism, telemetry and GPU/CUDA model execution.

### Data Scientist, Odds Notifier

Feb 2024 to Jul 2025 · Norway · Remote

- Python pipelines to collect, clean and analyze customer-support and social data from heterogeneous sources.
- Machine Learning and LLMs for message classification, prioritization, recurring-theme discovery and response automation.
- Power BI dashboards and reports for feedback, operational performance and product-improvement opportunities.
- Support for chatbot, content localization and testing of new features.

## Projects

### Supermarket Prices on iFood (personal project, featured)

Personal data-science project, rebuilt from scratch from an older version, covering the full path from scraping to a website. It collects public supermarket prices from the iFood website across the 27 Brazilian state capitals with a Python collector (about 1 request per second, no login, session from a real browser; challenges are never bypassed). Data is stored in PostgreSQL with a time-series schema, where each collection is a run. Outliers are cleaned with a MAD (median absolute deviation) rule. Comparisons between states and supermarket chains use the same items, and results are cross-checked against the DIEESE cesta básica and IPCA (IBGE). A Jupyter notebook documents the analysis, and a static pt-BR website with 8 pages shows maps, tables and charts (HTML, CSS, vanilla JavaScript, Chart.js, SVG map, light and dark themes). Results: 161k clean prices collected; 1,179 supermarkets across 27 capitals.

Technologies: Python, httpx, PostgreSQL, pandas, Jupyter, web scraping, Chart.js, JavaScript.

Demo: https://quanto-custa-o-mercado.vercel.app

### Production Demand Forecasting (professional work)

Demand-forecasting solutions for sales and operations planning (S&OP), with temporal validation, per-series selection, exogenous variables, multiple metrics and backtesting. Results: at least +10% accuracy (in percentage points) over the forecast used by multiple clients; model processing time reduced from about 8h to 2h with GPU/CUDA and pipeline optimization.

Technologies: Python, statistical forecasting, intermittent demand, gradient boosting, deep learning, foundation models, model ensembles, GPU/CUDA.

### External Data for Forecasting (professional work)

External data, calendars and hierarchies applied to forecasting. The workflow controls temporal availability, quality and persistence to prevent data leakage and keep backtesting comparisons consistent.

Technologies: Python, pandas, SQL, PostgreSQL, ETL/ELT, external APIs, data quality.

### Multiple (personal project)

Windows application built with C# and .NET to share mouse, keyboard, clipboard and audio across computers on the same network. Focus: low latency, UDP/TCP communication and automatic reconnection.

Technologies: C#, .NET 10, WinForms, UDP/TCP, Windows API, NAudio.

## Specialties

- Demand forecasting: statistical models, gradient boosting, neural networks, foundation models and ensembles.
- Validation and metrics: multiple metrics, temporal backtesting, per-series selection, leakage prevention, bias and uncertainty.
- Exogenous data: multiple exogenous datasets, calendars and hierarchies, with time-aware availability.
- Analytics product: indicators, intervals and analytical interfaces supporting planning decisions.
- End-to-end engineering: Python, C#/.NET, PostgreSQL and React/TypeScript from pipeline to interface.

## Forecasting models

- Statistical: AutoARIMA, AutoETS, Dynamic Optimized Theta.
- Intermittent demand: Croston, ADIDA, IMAPA, TSB.
- Machine learning: LightGBM, XGBoost, with lag, calendar, hierarchy and exogenous features.
- Deep learning: NHITS, NBEATSx, DeepAR, TFT, BiTCN.
- Time-series foundation models: Chronos, TimesFM, Toto.
- Ensembles and automatic best-model selection per series.

## Evaluation

- Metrics: bias, MAE, MAPE, MASE, sMAPE, WMAPE, scaled CRPS, SPEC and prediction intervals.
- Backtesting: holdout, expanding windows and sliding windows.
- Leakage prevention: exogenous signals are used only when they would have been available at forecast time.

## Tech stack

- Languages, backend and product: Python, SQL, C#, .NET, ASP.NET Core, EF Core, PostgreSQL, TypeScript, React, Apache ECharts, Power BI, Git.
- Data and forecasting ecosystem: StatsForecast, MLForecast and NeuralForecast (Nixtla), sktime, pandas, scikit-learn, PyTorch, Ray, NVIDIA CUDA, NumPy.

## Education

- Undergraduate Degree in Data Science, Centro Universitário Internacional (UNINTER), Nov 2023 to present. Focus: statistics and probability, machine learning, databases and SQL, data engineering and visualization.
- Data Science Specialization, Johns Hopkins University (Coursera), 2025 to 2026, completed. Covers R programming, statistical inference, regression models and practical machine learning.

## Certifications

- Claude Certified Developer. Anthropic, issued 2026-10-04, valid until 2027-10-04. Verify: https://www.credly.com/badges/ed806f2a-da35-485d-a8e0-d5d228b6968c
- Machine Learning Specialization. DeepLearning.AI and Stanford Online, 2025-04-16. Verify: https://coursera.org/verify/specialization/5J4V3DSKOX9W
- Supply Chain Management Specialization. Rutgers University on Coursera. Program: https://www.coursera.org/specializations/supply-chain-management
- Data Science Math Skills. Duke University, 2025-03-19. Verify: https://coursera.org/verify/H1UTPG2MKZJS
- Google Data Analytics Professional Certificate. Google, 2025-02-03. Verify: https://coursera.org/verify/professional-cert/Y4II56I38AR2
- CEFR C1 Advanced English. British Council (EnglishScore), Dec 2024.
- Microsoft Power BI Data Analyst Professional Certificate. Microsoft, 2025-02-21. Verify: https://www.coursera.org/account/accomplishments/professional-cert/GJHNN5B28NRB
- Python for Data Science, AI & Development. IBM, 2025-02-15. Verify: https://coursera.org/verify/8X0Y3VPS9GFK
- CS250: Python for Data Science. Saylor University, 2025-01-25. Verify: https://learn.saylor.org/admin/tool/certificate/index.php?code=5122948400TM
- AWS Cloud Technical Essentials. Amazon Web Services, 2025-02-22. Verify: https://coursera.org/verify/1AIZVC9OOX16
- Assess for Success: Marketing Analytics and Measurement. Google, 2025-02-20. Verify: https://coursera.org/verify/TXBHKWN7VR9Y
- Google AI Essentials. Google, 2025-01-24. Verify: https://coursera.org/verify/XCQWI4NJKV5V

## Languages

- Brazilian Portuguese (native).
- English (advanced, CEFR C1).

## Contact

Open to conversations about data science roles and projects. Email contato@theylor.dev or connect on LinkedIn: https://www.linkedin.com/in/theylor921