Médecins sans Frontières

Accurate demand forecasting is crucial for humanitarian equipment suppliers, such as Médecins Sans Frontières (MSF) Supply, to optimally allocate their resources to save lives. This project shows that with two models, based on ARIMA and the Gaussian Process Regression, forecasting of demand data can be improved compared to the model currently used by MSF. Additionally, important insights to forecastability of items as well as patterns in the ordering behavior are provided.

For more info, see the report and for the code Gitlab.

Students: Kathrin Durizzo, Frithiof Ekström, Carlos Garcia Meixide, Jonathan Koch

Mentor: Yevgeniy Ilyin

Helvetas

Helvetas manages around 300 projects per year in 30 different countries. For each project, a standardized report is manually generated which is required to be interpreted (in terms of progress, indicators, results, summary, etc). A natural language processing algorithm was developed which not only translates the reports, extracts and consolidates the relevant information, but also classifies the projects and understands the main topics, challenges and trends being tackled, enabling Helvetas management team to better communicate their impact, allocate resources and plan for the upcoming year.

For more info, see the report and for the code Gitlab.

Students: Catalina Dragusin, David Simon Tetruashvili, Jackson Stanhope, Tom Haidinger

Mentor: Marco Mancini

Impact Initiatives

Assessment of situational, demographic and infrastructural information in crisis regions is vital for organizing and executing the humanitarian response for local refugee communities. When key informants (KIs) in the region are difficult to reach and information becomes sparse, information reliability becomes paramount. In close collaboration with IMPACT initiatives, data collected throughout a study of KIs – local community leaders and experts from selected regions in Niger, Uganda, Afghanistan and Jordan – were analyzed with regard to each KI’s reliability. The data consisted of pairs of questions about social or infrastructural properties of their community and the KIs’ answers, respectively. After extensive data engineering, the driving features leading to high KI-reliability were investigated using tree-based regression models and explainable AI methods.

For more info, see the report and for the code Gitlab.

Students: Samyak Shah, Ayoung Song, Claus Wirnsperger, Feichi Lu

Mentor: Nima Riahi

Rega

In Switzerland, people hike a lot and in a normal year 20,000 people are injured and almost 40 incidents per year are fatal. That’s why Rega, the Swiss air rescue team, is constantly striving to make its rescue service faster and more reliable. Based on a patient’s location, the team developed a Lasso model, trained on previous missions, that predicts how long it would take each helicopter to reach the destination and then sends the fastest one.

For more info, see the report and for the code Gitlab.

Students: Colin Kälin, Rajiv Manichand, Elia Saquand, Hugues Sibille

Mentor: Stephan Artmann

Gesellschaft für Internationale Zusammenarbeit (ProSoil)

This project aims to support the GIZ’s mission to increase adoption of sustainable farming practices in western Kenya. The GIZ provided multiple farming Datasets out of which the team extracted insights and value. The team has developed a modular pipeline for data preprocessing and model training, including enrichment with publicly available geographical data. Furthermore, analysis of the data was conducted when possible, and recommendations on data acquisition were formulated to alleviate the issues the team encountered in the future.

For more info, see the report and for the code Gitlab.

Students: Antoine Basseto, Oscar Pitcho, Nando Metzger, Afshan Anam Saeed

Mentor: Lionel Trébuchon

Gesellschaft für Internationale Zusammenarbeit (Data Lab)

Tracking and tracing of the Nationally Determined policy implementations by countries is currently time-consuming and not systematically done, hampering accountability and sustainable progress. Therefore, GIZ requires a tool to quickly and automatically screen large amounts of documents (legal texts, speeches, tweets, …) for finding and contextualizing actual implementations of a nation’s commitments and policy priorities.  The approach developed consists of a combination of search tools, including a deep analysis tool and a faster simple keyword matching method. This allows automating the cumbersome task of manually screening dozens of documents regardless of the traced policy.

For more info, see the report and for the code Gitlab.

Students: Emily Robitschek, Raphael Sgier, Jonathan Doorn, Paul Türtscher

Mentor: Fran Peric