AI Early Warning Systems

The suite splits disaster forecasting into three separate, hazard-specific models rather than one generic predictor, because storms, droughts and floods behave on different timescales and respond to different signals. Tropical storm formation is framed as a short-horizon classification problem: a logistic regression model estimates the probability that a storm will form within the next seven days from current atmospheric and oceanic readings, built with Python, pandas and scikit-learn, with seaborn used to inspect feature relationships during development. Drought risk is a slower-moving, sequential problem, so it is modelled with an LSTM (a recurrent neural network built for time series) trained on environmental data in TensorFlow/Keras, with a Flask service exposing the trained model. Flood risk sits in between: it is scored with a custom multilayer perceptron built directly on TensorFlow's low-level API, combining meteorological and hydrological factors — rainfall, river levels, soil saturation — into a single flood-probability estimate.

Feeding these models is an ETL/ELT layer in Python, SQL and Spark that ingests from public sources including GIRI, EM-DAT, the World Bank and OECD CRS, standardised through PostgreSQL data marts and star-schema models so the same feature set can serve prediction, reporting and ad-hoc analysis without re-deriving it each time. A global data-governance taxonomy spanning 150+ countries keeps hazard, population and accessibility features consistent across regions, so a drought signal from one country's agricultural ministry and a flood gauge reading from another line up on the same schema rather than needing bespoke handling per source.

The financing side of the work is a live Power BI observatory, the UNDRR Disaster Risk Reduction Financing Observatory, built in collaboration with UNDRR itself rather than as a portfolio demo. It tracks five years (2019–2023) of official development assistance disbursements tagged under the OECD's DRR marker classification, turning donor and recipient-level aid flows into something a preparedness planner can actually query. An overview page filters by donor, year, country, sector, purpose and flow type, surfacing principal-versus-significant DRR disbursement totals, top donors and recipients, a recipient world map, and regional and sectoral breakdowns; a project-details page drills into individual projects by DRR/CCA objective and purpose code, with a full project-level table of donor, country, DRR status and total ODA.

Model outputs and the financing data converge into the same decision-support layer: instead of a forecast sitting in a notebook, the predictive models and the funding-gap analysis are built to answer the same question from two angles — where is the risk, and is the money going where the risk is. That framing, and the early-warning models behind it, won an international award and were presented at the Global Platform for Disaster Risk Reduction (GP2025) in Geneva, a forum mandated by the UN General Assembly, and featured in its official virtual exhibition.