Meraldo Antonio

My core quantitative work is in time series forecasting and nowcasting for macroeconomic and financial market variables — inflation, GDP growth, FX rates, credit spreads. I work with classical statistical methods (ARIMA, VAR), gradient boosting, and neural approaches, always with explainability as a first-class constraint. Beyond training, I deploy these models end-to-end: interactive dashboards in React with D3 visualizations, APIs in Python, all running on AWS. I contribute to sktime and skpro, the open-source Python libraries for probabilistic forecasting, which keeps me close to the research frontier.

Python • PyTorch • scikit-learn • React • D3 • AWS

My work involves mapping manual workflows and designing AI-augmented alternatives that keep humans in control.

Since early 2026, my focus has shifted toward implementing agentic AI across DBS Finance — moving the department from manual, spreadsheet-driven workflows to AI-augmented processes. This means less model-building and more change work: mapping existing processes, designing human-in-the-loop automation, getting buy-in from stakeholders, and building the tooling that makes new workflows actually stick. The organizational problems are just as hard as the technical ones.

Process Design • Workflow Automation • Stakeholder Management • AI Implementation

Designing RAG systems is part of what I do. This diagram shows a simple RAG system that uses both embedding-based and BM-25-based retrieval.

I build agent systems and automation pipelines using LangChain, LangGraph, and Claude's API — connecting language models to internal data through RAG pipelines, MCP tooling, and structured tool use. For faster iteration or no-code orchestration, I reach for n8n. The interesting problems are almost never the LLM itself: they're retrieval quality, tool design, failure modes, and deciding when not to use an agent at all.

Claude API • LangChain • LangGraph • n8n • MCP • RAG