Development Data Group

We leverage generative models to improve and scale metadata generation and refinement, including:

• Automated rewriting of indicator names and definitions for clarity and consistency.
• LLM-based validation of metadata completeness, coherence, and alignment with global standards.
• Multi-agent systems for iterative metadata improvement and enrichment.

We combine statistical methods, machine learning, and frontier AI models to detect, interpret, and contextualize irregularities in development data. This includes:

• Automated detection of anomalies using a hybrid approach that integrates rule-based checks, classical ML, and multimodal LLMs.
• Use of large language models to generate human-readable explanations for anomalies—especially when linked to identifiable causes such as conflict, economic shocks, or climate events.
• Exploration of AI agents that flag anomalies in near real-time and help data producers triage and investigate irregular data patterns.

We design and deploy AI-powered systems to enhance the discoverability of development datasets and indicators. This includes:

• Semantic search and recommendation systems for survey microdata, indicators, and documents.
• Retrieval-augmented generation (RAG) for better user interaction with statistical content.
• AI agents that assist users in refining their queries and identifying relevant data assets.

Explore how semantic- and keyword-based searches differ in the widget below.

We develop AI systems that track how development data is used in research, policy, and public discourse. This includes:

• Fine-tuned language models that identify and classify mentions of datasets in research publications and project documents.
• A data citation taxonomy that distinguishes between named, unnamed, and superficial data references to evaluate the depth of use.
• Systems to quantify the visibility, reuse, and policy relevance of datasets across countries, sectors, and languages—laying the groundwork for open, citation-based metrics for data impact.

We focus on bridging the computational and socio-linguistic gaps in AI to ensure that tools are effective, fair, and usable across all development contexts, helping ensure that low-income countries are not left behind. This includes:

• Developing methods to evaluate and surface biases in AI systems, particularly those affecting low-resource languages and underrepresented populations.
• Training compact, low-resource embedding models to enable efficient semantic search, classification, and retrieval in bandwidth- or compute-constrained settings.
• Advising on and implementing practical AI strategies for countries with limited data coverage, including guidance on leveraging AI with non-traditional or incomplete datasets.