$150,000 in grants for research on public-goods funding.

1. 01Optimising Anti-Poaching Investments in Protected Areas Under Increasing Resource ConstraintsCharles A. Emogor, Anil Madhavapeddy, Rob Fletcher, Lingkai Kong, Inaoyom ImongUniversity of Cambridge · Harvard University · Wildlife Conservation Society, Nigeria🇬🇧 🇺🇸 🇳🇬

   Abstract

   Despite over $3.6 billion spent globally on anti-poaching efforts since 2010, illegal wildlife exploitation persists, driven by local and international demand and constrained resources. Effective deployment of ranger patrols is critical, yet current strategies rely primarily on patrol-collected data, which are biased due to uneven coverage and imperfect detection. This project addresses these limitations by integrating ranger and hunter-collected data to generate robust, bias-corrected maps of poaching in three ecologically significant protected areas (PAs) in southeast Nigeria: Cross River National Park (Oban and Okwangwo divisions), Afi Mountains Wildlife Sanctuary, and Mbe Mountains Community Forest. Ranger patrols (2016–2025) and hunter GPS data (collected from 120 hunters across eight communities since June 2025) provide independent, complementary records of poaching activity.

   We will first model poaching patterns using random forests on hunter data to generate continuous spatio-temporal poaching intensity surfaces independent of ranger detection. Ranger detections will then be modelled using a diffusion-based framework that explicitly handles imperfect detection and predicted ‘true’ poaching, allowing systematic under-detection to be quantified. The resulting bias-correction layer will be applied to full-site diffusion models, producing corrected poaching risk surfaces that reflect true poaching intensity.

   Building on these forecasts, we will employ reinforcement learning to design adaptive, cost-effective patrol strategies that maximise deterrence under resource constraints, including patrol size, frequency, and routing. By collaborating with the Wildlife Conservation Society, outputs will be directly integrated into patrol planning, enabling rapid translation of research into practice, strengthening local capacity, and improving the effectiveness of anti-poaching interventions across the landscape.

2. 02Deterministic Governance and Traceability Layer for AI-Assisted Public Goods FundingSungwoo JungKorea University🇰🇷

   Abstract

   Quadratic Funding (QF) mechanisms are mathematically deterministic, yet the pipelines that execute them are not. Real-world public goods allocation workflows involve preprocessing steps—Sybil filtering, deduplication, normalization, and parameter tuning—implemented as ad-hoc scripts that fail to capture intermediate states and implicit parameters. As AI tools enter funding workflows, non-deterministic model outputs further blur the boundary between “AI suggestion” and committed computation.

   We propose a Deterministic Governance and Traceability Layer that wraps existing allocation pipelines without replacing them. Built on the Manifesto deterministic state framework, the system captures every input, parameter change, and intermediate state as content-addressable snapshots linked in an append-only lineage graph. Each change is governed: Actors (human operators or AI systems) submit Proposals that require Authority approval before execution. AI tools act as untrusted proposers whose influence is explicitly recorded and auditable.

   We evaluate the approach on public Gitcoin Grants round data, comparing replay reproducibility, audit query response time, and AI influence visibility against a standard scripted baseline. The result is a practical governance wrapper that enables reproducible, verifiable allocation workflows and structured post-round impact analysis—without requiring infrastructure rewrites.

3. 03Bread: Grant Access for Women-Focused Nonprofits Through AI-Powered Discovery and Proposal GenerationRopafadzo Shumba, Rose ShumbaUniversity of Maryland Global Campus · Bowie State University🇺🇸

   Abstract

   Less than 2% of US charitable giving reaches women's and girls' organizations, yet $110 billion in foundation grants are distributed annually. Small nonprofits serving women lack the capacity to find and apply for grants — a single application requires 80-200 hours, and most organizations have only 1-2 staff managing the entire process. Existing grant technology assumes users already have grant-writing experience and staff, leaving grassroots organizations behind.

   Bread is an AI platform that discovers grants and generates proposals specifically for small, women-focused nonprofits with no grant-writing background. Built on the Claude API with retrieval-augmented generation, Bread performs three functions: (1) grant matching using semantic similarity across federal, foundation, and corporate funding sources, with IRS 990 filings parsed to identify funders with historical giving to women-focused causes; (2) AI proposal generation that reduces 80-200 hours of work to 2-3 hours of review; and (3) sector-specific intelligence incorporating gender-focused funder language and frameworks.

   The research methodology proceeds in four phases: building an open-source dataset of 2,000+ grant opportunities, developing and evaluating three matching approaches (BM25, dense retrieval, hybrid), building a RAG pipeline using Claude Opus 4.6, and deploying a pilot study with 15-20 small women-focused nonprofits. Quality is assessed through blind evaluation by independent grant writers comparing AI-generated drafts to human-written proposals.

   Deliverables include an open-source grant dataset, a working prototype, a comparative research report, and all source code under MIT license. The primary research question: can an AI platform increase both the volume and quality of grant applications from small women-focused nonprofits?

4. 04Mapping the Mismatch: AI-Assisted Analysis of Public Goods Funding Priorities and Local NeedsSteven W. Witt, Margarita Gómez, Michelle QuiñonesUniversity of Illinois Urbana-Champaign · Southern Voice🇺🇸

   Abstract

   Mapping the Mismatch: AI-Assisted Analysis of Public Goods Funding Priorities and Local Needs is a collaborative initiative between the University of Illinois’ International and Area Studies Library (IASL), the Center for Global Studies (CGS), and the Southern Voice Think Tank Network. The project addresses a structural gap in global public-goods funding: the absence of scalable, transparent tools to assess whether donor priorities align with needs articulated by research and policy institutions in the Global South.

   The team will build two complementary corpora: (1) publicly available donor strategies, calls, and funded project descriptions, and (2) approximately 350 publications from Southern Voice’s Knowledge Hub and Climate Change Database representing institutions across Africa, Asia, and Latin America. Using multilingual natural language processing methods—including document embeddings, topic modeling, clustering, and similarity analysis—the project will identify thematic convergence and divergence between funding discourse and Southern-led research agendas.

   Designed as a decision-support and accountability tool rather than an automated allocation system, the project will produce a comparative gap analysis, a policy brief for climate and development stakeholders, and a documented, reusable methodology to strengthen transparency, equity, and evidence-informed dialogue in global public-goods governance.

5. 05Mental Health in Public Indian Universities: Learning from Chatbot-Generated Behavioral and Engagement DataAnwesha Bhattacharya, Bhavya Srivastava, Akanksha Aggarwal, Suha S. Dewan, Abhilasha MansataHarvard University · The World Bank🇺🇸 🇮🇳

   Abstract

   Mental health is a severely underfunded public good in low- and middle-income countries (LMICs), where governments and nonprofits must allocate scarce resources without real-time information on need, engagement, or stigma. With a 95% treatment gap in mental health care, this challenge is acute in India. In surveys of around 5,000 students at a large public university, distress is high: almost 50% students screen positive for symptoms of depression and anxiety. Yet, only 6% report past-year counseling. Gender inequities are stark: female students report substantially higher distress and stigma-related constraints, while male students face norms discouraging vulnerability. Demand-side frictions, such as fear of initiating care, uncertainty, stigma, and perceived judgment, exacerbate supply-side shortages and make targeting scarce mental-health resources challenging.

   This proposal introduces MindMitra, an open-source, non-clinical mental health chatbot fine-tuned on culturally grounded synthetic therapy dialogues using LLM-based methods. Co-designed with Indian therapists, psychologists, economists, and computer scientists, MindMitra offers a private, bilingual, low-friction entry point into care. It functions as a bridge to help students articulate concerns, reduce stigma, and build confidence to seek professional support.

   The proposal’s contribution to public goods funding is twofold. First, MindMitra generates real-time, privacy-preserving engagement that can help universities identify unmet needs. Second, through a randomized controlled trial, we will produce causal evidence on which populations benefit most and which engagement patterns predict help-seeking through a layered experimental design. All tools, datasets, and model pipelines will be released as open source.

6. 06Robust Randomization: Smooth Partial Lotteries for Grant FundingAlexander Koujianos Goldberg, Giulia Fanti, Nihar B. ShahCarnegie Mellon University🇺🇸

   Abstract

   Traditional grant review aims to select high-quality candidates from reviewer evaluations. These processes face criticism for error-prone reviews, bias toward safer projects, and substantial time spent deliberating near the funding line. Indeed, grant peer review is highly resource-intensive, costing up to 20-35% of total budgets. In response, scientific funders around the world have introduced randomization into decisions. These funders run “partial lotteries” that randomly allocate acceptances based on peer reviews of proposal quality. Lotteries are intended to soften score cutoffs and reduce time spent distinguishing between similar proposals. However, recent work, including ours, shows that existing lotteries can be unstable. Small changes in review scores induce large jumps in proposals’ selection probabilities. Minor factors like reviewer fatigue could significantly impact outcomes, creating arbitrariness. Further, reviewers may still spend considerable time agonizing over small quality differences.

   In this project, we will develop Smooth Partial Lotteries, a new class of randomized mechanisms for grant allocation in which selection probabilities change gradually with review scores. We design these lotteries to balance quality (funding the highest-rated proposals) with smoothness (defined as a Lipschitz condition on the mapping from reviews to selection probabilities). Our project will (1) develop mechanisms that achieve provably optimal trade-offs between smoothness and quality, (2) release an open-source Python implementation, and (3) validate methods on real grant and peer-review datasets. We will leverage existing collaborations with funding agencies to position Smooth Partial Lotteries as a practical upgrade to existing lotteries, reducing reliance on human deliberation through AI-powered grant allocation.

7. 07A Data-Driven Decision Intelligence Approach to Enhance Community and Environmental Health Resilience in KenyaHaipeng Chen, John Sankok, Carrie Dolan, Erica Garroutte, Julius Odhiambo, Nnamdi OsakweCollege of William & Mary · Community Health Partners, Kenya🇺🇸 🇰🇪

   Abstract

   Climate change is reshaping weather patterns, degrading ecosystems, and altering natural resource availability, with significant consequences for human health. Nomadic communities who rely directly on natural resources—such as the Maasai in Narok County, Kenya—are especially vulnerable, facing both heightened climate-driven stress and limited access to essential health services. Understanding how environmental conditions interact with human health outcomes is crucial for improving the accessibility and allocation of healthcare resources, yet this remains difficult due to the complexity and cross-sectoral data required.

   This project applies a data-driven decision intelligence (D3i) framework to integrate environmental, ecological, and public-health datasets in order to identify relationships that influence health vulnerabilities in nomadic communities. Working with Kenyan partners, we will (1) collect and curate interdisciplinary datasets needed for prediction and decision support; (2) develop and validate predictive models capable of forecasting public-health needs under changing environmental conditions; and (3) design and evaluate a decision-making method that allocates limited healthcare resources more effectively and equitably. The resulting tools aim to strengthen the resilience of highly vulnerable populations in Kenya by enabling informed, data-driven responses to climate-related health risks.

8. 08Equity-Aware Explainable AI for Funding Allocation in Digital Public Goods: Addressing Hate Speech in Latin AmericaFrancielle Vargas, Jackson TragerUniversidad de Chile · University of Southern California🇨🇱 🇺🇸

   Abstract

   Digital public goods aimed at mitigating hate speech and harmful online content increasingly receive public and philanthropic funding. However, funding allocation processes remain largely qualitative, English-centric, and insufficiently equipped to evaluate region-specific risks and impact potential, particularly in underrepresented ecosystems such as Latin America. As a result, initiatives grounded in local linguistic, cultural, and sociopolitical realities may be undervalued, while unintended amplification risks remain insufficiently assessed. This paper proposes an equity-aware, explainable NLP-based decision-support framework to assist funding allocation for digital public goods addressing hate speech in Latin America. The system integrates Retrieval-Augmented Generation (RAG), structured harm modeling, and rationale-guided explanations to assess: (1) projected harm mitigation capacity, (2) amplification and governance risks, and (3) distributional fairness across linguistic and regional communities. A central contribution is the introduction of fairness-aware funding simulations that evaluate how AI-assisted allocation may differentially impact Portuguese- and Spanish-speaking ecosystems, marginalized groups, and locally developed moderation infrastructures. The framework is designed as a human-in-the-loop tool, prioritizing transparency, accountability, and contextual sensitivity. By embedding explainable and equity-centered AI into funding governance processes, this work aims to strengthen the effectiveness and fairness of digital public goods funding in Latin America while offering a model adaptable to other Global South contexts.

9. 09Cross-Lingual Impact Verification (CLIV) for Public Goods FundingAye Hninn KhineKing Mongkut's University of Technology Thonburi🇹🇭

   Abstract

   This project addresses this by developing a specialized AI framework for Cross-Lingual Impact Verification (CLIV). Using a multi-agent Large Language Model (LLM) architecture, we automate the extraction of event-level impact data—such as aid delivery, infrastructure milestones, and crisis severity—from Burmese and regional news sources. By fine-tuning publicly available SEA-centric models (e.g., SEA-LION), we bridge the gap between high-level grant objectives and on-the-ground reality. This project transforms local journalism into a decentralized verification layer, making funding more transparent and accountable. Our 6-month timeline focuses on building a prototype that generates an "Impact Score" for funded projects by cross-referencing grant claims against real-time news extracted from low-resource linguistic environments. This research directly supports the Fast Grants' mission to improve governance and community coordination in the Global South.

10. 10An AI-Powered Funding Database for the African Tech EcosystemAya Salama, Sabrina AmroucheMicrosoft · AfriClimate AI🇪🇬 🇿🇦

    Abstract

    Identifying relevant non-equity funding opportunities remains a significant challenge for the African technology community due to fragmented sources and limited visibility. This proposal develops a centralised, AI-powered database that curates funding opportunities with structured metadata including descriptions, deadlines, and eligibility criteria. Motivated by the Deep Learning Indaba Ideathon's manual approach to sharing opportunities, and validated by survey data showing 62\% of respondents missed opportunities in the past year, the project employs modern language models and information extraction to automatically populate and maintain the database. The work proceeds through directory curation, AI-assisted extraction, monitoring implementation, and pilot validation with Ideathon teams. The resulting infrastructure will support future recommendation systems for matching projects to grants, directly advancing AI4PG's goal of improving public goods funding allocation.

11. 11Scaling Citizen Input Through AI Summarization in Timor-LesteJosé Ramón Enríquez, Julio S. Solís Arce, Jiaxin Pei, Lula Chen, Lily Tsai, Alex Pentland, Venancio Pinto, Elizabeth BaptistaStanford University · Harvard University · MIT · TIC Timor I.P.🇺🇸 🇹🇱

    Abstract

    Can AI-generated summaries of collective preferences improve citizen input quality in fragile democracies? We partner with Timor-Leste's Information Technology Agency to conduct a survey experiment gathering public opinions on the Digital Transformation Strategy (2025–2036) in the capital city Dili. Participants engage with an AI conversational agent that presents summaries of in-group and out-group preferences. Our three-arm design tests whether AI summarization of collective views improves argumentation quality: (1) Control with basic policy text; (2) Factual synthesis presenting neutral summary of what different groups prefer; (3) Trade-off summary adding explicit benefits and costs each group emphasizes. Timor-Leste --- a post-conflict state with limited government capacity and linguistic diversity --- exemplifies contexts where traditional consultation risks amplifying partisan, geographical, and generational divisions. AI summarization of collective preferences offers a scalable approach: exposing citizens to diverse viewpoints through neutral conversation. We measure three outcomes: (1) opinion quality and preference convergence; (2) value of information through willingness-to-pay; (3) respect for and willingness to support policy outcomes. Comparing arms identifies whether exposure to collective preferences alone improves deliberation or whether explicit trade-off framing is necessary. This pilot on digital policy will inform future applications to health and education priority-setting.

Reviewed & selected on OpenReview · AI4PG 2026

OpenReview → openreview.net/group?id=AI4PG/2026