Award-winning ML project
CrisisLens
CrisisLens is an AI-assisted command center for crisis prioritization that turns fragmented humanitarian signals into operational, country-level decisions. Our team won **1st Place** in the Databricks x United Nations Challenge at Hacklytics (Georgia Institute of Technology). **Core Capabilities:** • **Real-time 3D globe command center** for country-level risk, funding, and operational signals • **Databricks Genie integration** for natural-language querying and structured analytical outputs • **Simulation workflows** with multi-quarter projections and impact visualization • **ML forecasting** to estimate crisis trajectories before escalation • **Multimodal navigation** across pointer, pinch, hand-tracking, and voice commands **My Contributions:** • Cleaned and structured multi-country crisis datasets • Performed EDA to identify temporal and funding patterns • Ran SHAP analysis and feature engineering • Trained ensemble models (LightGBM, Random Forest, XGBoost, Gradient Boosting, Stacking) • Applied group- and time-aware cross-validation to prevent leakage • Built multi-horizon forecasting (6-month and 12-month+) • Implemented KNN peer mapping for cross-country similarity analysis • Integrated predictive outputs into the interactive visualization layer **Team:** Jakob Laise, Alexander Paolini, and Abduaziz Umarov **Built With:** Databricks, Next.js, React, TypeScript, Tailwind CSS, Three.js, Docker, NumPy, Pandas, PyTorch **Project Links:** • Project: https://lnkd.in/essqGkQZ • Live Demo Site: https://crisislens.paolini.dev • GitHub: https://github.com/Michael-RDev/CrisisLens • Video Demo: https://youtu.be/YD6H3emkDrc
Problem
Humanitarian teams must prioritize aid while crisis, funding, and operational signals are fragmented across countries and time.
My role
Cleaned and structured the datasets; performed EDA, feature engineering, and SHAP analysis; trained and validated the forecasting ensemble; built peer mapping; and integrated predictions into the visualization.
Approach
Fuse multi-country data with leakage-aware ensemble forecasting, SHAP analysis, KNN peer mapping, Databricks Genie, and an interactive 3D globe.
Outcome
Won 1st Place in the Databricks × United Nations Challenge at Hacklytics and delivered a working crisis-prioritization command center.
Implementation
- Built 6-month and 12-month-plus forecasting workflows with multiple ensemble models.
- Used SHAP for feature analysis and KNN for cross-country peer mapping.
- Connected model artifacts, Databricks analysis, simulations, and a Three.js interface.
Evaluation
- Applied group- and time-aware cross-validation.
- Compared LightGBM, Random Forest, XGBoost, Gradient Boosting, and stacking approaches.
- Reviewed feature drivers with SHAP rather than treating predictions as opaque scores.
Results
- Produced multi-horizon country forecasts and scenario outputs inside the live interface.
- Awarded 1st Place in the Databricks × United Nations Challenge at Hacklytics.
Constraints
- Heterogeneous country-level data needed a consistent schema.
- Time and group leakage had to be prevented during validation.
- Forecasts needed to remain understandable inside an interactive decision tool.
Tradeoffs
- The ensemble improves coverage at the cost of model and deployment complexity.
- Scenario outputs support decisions; they are not a substitute for field intelligence.
Next improvements
- Add calibrated uncertainty intervals and more explicit data-quality indicators.
- Validate prioritization workflows with humanitarian-domain users.