I don't just write code — I solve business problems. Here's how I approach every engagement.
Identify the real bottleneck — manual hours, inaccurate reports, missing signals. I map the current workflow and quantify the cost of inaction before touching a single line of code.
Real-world data is messy. I build robust pipelines that handle duplicates, nulls, encoding issues, and schema drift — so the analysis is always grounded in clean, trustworthy data.
I choose the right tool for the job — ML model, ETL pipeline, or Power BI dashboard. The goal is always the same: deliver a system that runs itself and improves over time.
Every output is validated against ground truth and backtested where possible. I document the measurable business impact — time saved, accuracy gained, revenue at risk identified.
Python 
C# / .NET 
JavaScript 
HTML & CSS 
Power BI (DAX) 
SQL (ETL) 
Excel (VBA) 
Machine Learning 
LangChain / RAG 
Streamlit Apps 
Pandas & NumPy Real activity. Public repos. Every project is open-source and verifiable.
"His ability to learn and apply acquired knowledge in new situations truly impressed me, predicting a promising professional career. José will surely be an asset to any company."
Financial modeling of automated workflow returns
Lessons learned from real projects — not tutorials.
Logistic regression is interpretable but breaks on non-linear interactions — which are everywhere in financial data. Random Forest handles correlated features gracefully, gives you feature importances for free, and rarely overfits with basic hyperparameter tuning. The Banker AI project validated this: 85% accuracy vs 71% for a tuned logistic baseline.
Everyone automates the calculation. Nobody automates the validation. The Financial Reporting project taught me that 70% of errors come from silent data-type mismatches between source files. I now always build a validation layer that halts execution and surfaces the exact row and column that failed — before writing a single output cell.
Building ChatCFO, I found that 512-token chunks with 15% overlap outperformed 1024-token chunks for financial PDFs. Larger chunks dilute the dense numerical context that embeddings need to stay precise. The sweet spot depends on document structure — tabular data needs tighter chunks than narrative prose.
In the E-Commerce project, a customer generating €5k/year in revenue appeared healthy — until RFM revealed they hadn't purchased in 9 months. Raw revenue is a lagging indicator. Recency is a leading one. That single insight unlocked a €96k revenue-at-risk identification that would have been invisible in a standard sales report.
I'm always open to discussing data-driven projects, automation workflows, or potential roles in Finance & Tech.
