Data Scientist & ML Engineer — Belo Horizonte, working remotely
Forecasting, applied ML and LLM systems for BMW Financial Services Canada, Johnson & Johnson LATAM and public health research — plus the warehouse modelling and orchestration underneath them. Before data, metallurgical engineering: a discipline of tolerance, specification and failure analysis. It shows up in how the work gets done.
Allocated to BMW Financial Services Canada and Johnson & Johnson LATAM, then to an internal commercial intelligence product.
BMW Financial Services Canada. Built a production pipeline processing 3,000+ monthly survey comments into 25+ structured attributes — 3,094 comments in ~25 minutes at 92% accuracy against human annotation, replacing partial manual review. Built a three-tier sales forecasting system across ~1,800 retailer-model combinations, orchestrated as a Step Functions state machine with weekly data refresh and monthly retraining promoted to production. Consolidated commercial finance data from on-premise SQL Server into a five-table warehouse asset, cutting query runtime ~60% on 15M+ row datasets.
Johnson & Johnson LATAM. Warehouse capacity forecasting across three distribution centres — XGBoost at SKU level, MAE 1.79 bins, WMAPE 6.2% at storage-type level, outperforming a moving-average baseline. Built a Streamlit what-if simulator for scenario planning, with SHAP explainability so any surprising forecast can be audited before a decision.
Internal product. Built the classification core of a commercial intelligence system reading sales call transcripts, with structured output validation, per-dimension justification and a pseudonymization pipeline measured as zero-cost in accuracy.
Longitudinal sentiment analysis of public reviews of Brazil's public health application, using a multilingual transformer pipeline. Published at EnANPAD 2025.
ETL from PostgreSQL with data integrity checks, optimised SQL extraction, and a Streamlit interface for exploring device performance across daily, monthly and yearly time scales.
Supported students through a data science bootcamp — Python, statistics, machine learning, deep learning and deployment fundamentals. Debugged student code live and gave structured feedback on projects.
MBA in Big Data, Descomplica (in progress) · Natural Language Processing, UFMG · Metallurgical Engineering, PUC Minas · Engineering exchange, Queensland University of Technology (CNPq scholarship)
A personal job assistant. Multi-source ingestion across four ATS APIs, three-layer deduplication, a validated state machine, idempotent scheduled runs.
Python · Supabase · GitHub Actions · Claude APIMulti-agent system for the CloudWalk technical challenge. Routing rules derived from exploratory analysis of 99 real messages, full test coverage, containerised.
LangGraph · Python · DockerRetrieval-augmented generation over apiculture domain content.
Python · RAG · embeddingsMeasure rather than assume. One variable at a time. Document what you decided not to do.
I came to data through metallurgical engineering, and before that, years as a tattoo artist. Both are disciplines of tolerance and irreversibility — you specify before you cut, and there is no undo. That is the through-line: the work I trust is the work that was measured, and the numbers I publish are the ones I can defend under a follow-up question.
Today I work across forecasting, NLP and LLM systems, and the pipelines that carry them. The part I find most interesting is not the model — it is what happens when the model is wrong and nothing raises an exception.
Currently open to remote roles and consulting through AIRA Labs, my applied AI studio. Based in Belo Horizonte, working with distributed teams.