Athena

Inspiration

Every year, the United Nations identifies hundreds of millions of people in need of life-saving humanitarian aid. Yet a hard truth persists. Aid is not always distributed where the need is greatest. It is often distributed where attention is highest. While some crises receive close to 100% of their requested funding, others are left behind. These so-called forgotten crises sometimes receive less than 30% of required funding, even when the severity of suffering is comparable.

What makes this gap more troubling is visibility. Today, UN decision-makers must sift through dozens of spreadsheets, fragmented datasets, and static PDF reports to uncover these mismatches. Critical signals are buried across tools and formats. In moments where speed and clarity matter most, decision-makers are effectively flying blind. We realized this gap was not a data problem, but a visibility and decision problem. Athena was built to surface overlooked crises clearly, explain why the gaps exist, and help decision-makers act with confidence.

What it does

Athena highlights mismatches between crisis severity and pooled fund coverage on a global map, making underfunded yet high-need regions immediately visible. It analyzes project-level data to flag unusually high or low beneficiary-to-budget ratios, enabling meaningful benchmarking rather than surface comparisons. Athena then explains the context behind these discrepancies, verifies claims using credible sources, and suggests data-backed actions. In urgent scenarios, it can notify relevant representatives through a multilingual voice agent so insights move beyond the screen and into action.

How we built it

We built Athena using public UN datasets covering crisis severity, funding coverage, and project-level indicators such as budgets, clusters, and beneficiary counts. At the core is an Isolation Tree model with depth 200 that detects anomalous beneficiary-to-budget ratios across projects. The model was trained, tested, and deployed on Databricks, then securely served through an endpoint.

These results are visualized on an interactive map and globe that surfaces regions of interest and funding gaps. We added an intelligence layer powered by Gemini that interprets model outputs, explains patterns concisely, verifies its reasoning using web search, and retrieves context through dataset-connected RAG and visualizations. A voice agent built with ElevenLabs was embedded into a phone line to deliver emergency notifications.

The frontend was built with JavaScript, HTML, and Tailwind CSS, the backend with Flask, Python, scikit-learn, NumPy, and pandas, deployed via Vercel, and made accessible to all users through Userway API integration.

Challenges we ran into

Our biggest challenge was deploying and serving the ML model on Databricks. Moving from a working notebook to a reliable, production-ready endpoint required navigating configuration issues, deployment failures, and tooling constraints. We contacted multiple Databricks representatives, tested several approaches, and iterated repeatedly under time pressure. The experience reinforced how challenging it can be to move from experimentation to production, even with a strong platform. We ultimately succeeded in deploying the model and serving it securely through an endpoint.

Accomplishments that we're proud of

We are proud to have deployed a production-ready ML model on Databricks within a hackathon timeframe. We are equally proud of building a system that goes beyond visualization to deliver decision-ready intelligence. Most importantly, Athena was designed with accessibility at its core, ensuring it can be used by people with vision, hearing, and learning challenges.

What we learned

We learned that strong impact does not require excessive complexity. A single well-chosen ML model, when paired with verification, context, and thoughtful UX, can outperform far more complex pipelines. We also learned how critical reliability and clarity are when building tools for high-stakes humanitarian decision-making.

What's next for Athena

Next, we plan to expand Athena with forecasting and early-warning capabilities so UN analysts can identify funding risks before crises escalate. We also aim to integrate additional humanitarian datasets and refine the recommendation engine. With further development, Athena can evolve into a real-time decision-support system that helps ensure no crisis is overlooked simply because it is less visible.