FluCast Outbreak Forecasting

Context & Goals

Built FluCast during a public health research initiative to help epidemiologists anticipate respiratory surges. Needed reliable ingestion from CDC ILINet APIs, configurable regional models, and dashboards for health departments.

Architecture

• Data ingestion: Scheduled ETLs pull CDC feeds, sanitize anomalies, and persist to PostgreSQL + S3 history buckets.
• Modeling: PyTorch LSTM ensemble with feature engineering for seasonality, weather, and mobility metrics. Azure ML orchestrates retrains.
• Serving: FastAPI inference service with caching and guardrails that expose forecasts and confidence bands to the Next.js frontend.
• Visualization: Next.js dashboard with map overlays, scenario comparisons, and PDF export for weekly reporting.

Challenges & Wins

• Automated monitoring detects concept drift and sends alerts if confidence intervals widen beyond thresholds.
• Implemented reproducible experiment tracking using MLflow, enabling quick audits of model changes.
• Added region-level overrides so public health analysts can tweak interventions and see projections instantly.

What I’d do next

• Expand to include wastewater and school absence data for better leading indicators.
• Introduce uncertainty-aware ensembles and quantile regression for richer decision support.