Armada
Multi-agent orchestration framework for coordinating AI coding teams through phased contracts with evidence gates, role dispatch, and content collections.
Researcher and builder World Bank
Rafael Macalaba
Research, engineering, and applied AI.
I’m exploring small language models for low-resource settings, including on-device inference in the browser. Other threads include search and reranking systems, synthetic data, and practical AI evaluation.
From research to working tools
Selected work
ai4data
World Bank toolkit for AI-assisted analysis of development data.
data360-mcp
Model Context Protocol server for the World Bank Data360 platform.
Ideas, methods, and field notes
Recent writing
armada: orchestrating AI agent teams with evidence gates
Introducing armada, an open-source tool that splits coding work into evidence-gated phases and dispatches specialist subagents.
Field Order Should Not Matter: Permutation-Invariant Embedding Model Fine-Tuning for Structured Metadata Retrieval
Permutation-invariant fine-tuning approach for embedding models applied to structured metadata retrieval.
AI for Monitoring and Classifying Data Used in Research Literature
Methods for AI-assisted monitoring of how datasets appear in published research, including classification of usage context.
Current focus
What I’m working through now
Lately I have been going deep on semantic search — embeddings, rerankers, the whole retrieval pipeline. The problem that keeps me up is not just finding relevant documents but understanding why they are relevant, which means getting comfortable with how dense retrievers collapse semantic space and where hybrid sparse/dense approaches break down.
Continue the conversation
Working on a data or AI problem?
Reach out about research, software, or collaboration.