ICTworks

In the complex world of international development, understanding what works and what doesn’t is crucial to maximizing the impact of foreign aid. Traditionally, this has meant either relying on the anecdotal evidence from your peers and personal experience, or wading through mountains of project evaluations to find actionable insights.

This is a daunting task given the sheer volume of data scattered across cumbersome databases or buried in the USAID Development Experience Clearinghouse (DEC), where too many good project reports are buried in PDF graveyards.

DELLM: GenAI Answers from the DEC

Happily, this time-consuming discovery process now moves at the speed of AI: introducing the Development Evidence Large Learning Model (DELLM), a Generative AI solution fine-tuned to the unique domain-specific needs of the international development sector.

DELLM is engineered to conduct an exhaustive analysis and interpretation of hundreds of thousands of project evaluations in the DEC and to triangulate findings with various other data sources from academic articles to local data sets. This GenAI capability enables DELLM to amalgamate and deliver cohesive insights from a broad spectrum of data, markedly enriching our comprehension and assessment of development evidence.

The result is a tool that eclipses traditional manual coding in speed—outpacing what would take a team of four researchers 40 years to accomplish—and in depth, providing reliable access to all evidence within the DEC and beyond.

USAID DEC Insights with DELLM

USAID staff can use DELLM to streamline processes and cut red tape by leveraging the DEC today. Read our case study for how DELLM helped USAID/Tajikistan automate PPR narrative generation, reclaiming over 1,250 annual staff hours and mitigating the need for extended work days to adhere to critical deadlines.

Implementing partners can also use DELLM. Here are two ways to find DEC insights for better program design:

1. Identify USAID Funding Trends

A team wanted to know the trends in USAID interventions over time to design their agriculture strategy. DELLM categorized every agriculture intervention over the past 60 years and charted their popularity over time. The team discovered that the research and development intervention peaked in usage in 1987 when it was used in 50% of projects. Then a dramatic decline happened and only 5% of agriculture projects use R&D today.

Why? Further research found that the Bumpers Amendment curtailed R&D on agricultural commodities. More importantly, agriculture programs didn’t use robust data in Congressional budget justifications, unlike health, which started using strong impact evidence and saw its budget increase dramatically.

2. Win USAID Contracts

A different business development team was bidding on an environmental conservation project. They queried the entire DEC to scrutinize their competitors’ past performance.

DELLM helped them identify three competitors with strong past performance but weak long-term sustainability. For example, one competitors’ reforestation project used non-native species, disrupting the local ecosystem. DELLM also proved the team’s clear commitment to sustainability and local community integration.

They found their win theme! They highlighted their commitment to lasting impact and community empowerment, and backed it with compelling evidence, which won them the contract. Insight they would’ve missed without DELLM.

How DELLM Delivers Accurate Results

The core of DELLM is a huge manually-coded dataset developed by a team of researchers and data scientists over five years and peer-reviewed by USAID and the United Nations.

This dataset, comprising hundreds of thousands of labeled excerpts from carefully curated sources, is what enables DELLM to navigate the intricate terminologies and contexts of development work just as development practitioners would.

For example, have you ever tried to ask ChatGPT questions like, “which interventions are most effective at empowering women entrepreneurs in Pakistan?” Try it now! You’ll likely receive a uselessly generic response sourced from the vast expanse of the Internet and influenced by the gender biases ingrained in that training data.

Using public GenAI tools like ChatGPT, Copilot, or Gemini to design your programming can undermine effective efforts to increase aid effectiveness, and increases risks like perpetuating misinformation, something the World Economic Forum tagged as the most important global risk in 2024.

DELLM’s domain-specific training delivers equitable and precise understanding, mitigating the risk of perpetuating biases and enhancing the effectiveness of international development efforts. We also compared DELLM with ChatGPT and other GenAI solutions and were blown away by DELLM’s remarkable capability to accurately classify interventions. We found that:

• DELLM F1 scores outperform ChatGPT, which means DELLM does a better job of understanding and responding accurately to user requests.
• DELLM recall rates exceed 85% allowing its use for decision-making, while ChatGPT usually scores around 65% in academic tests.

You Can Use DELLM Now

DELLM is available now for leading international development firms. Please contact us to get started using DELLM to find insights in the DEC or your own firm’s data and reporting.

USAID, DT Global, Chemonics, and other leading organizations are pioneering DELLM integration into their proposal processes and tailoring it to their unique expertise for enhanced evidence discovery and analysis.

Don’t miss out – this is the dawn of a new era in foreign aid, propelled by GenAI’s transformative impact on efficacy and insight.

By Lindsey Moore CEO of DevelopMetrics