Overview#
Designed and implemented a real-time river level forecasting system integrating telemetry ingestion, preprocessing pipelines, machine learning models, and automated deployment.
Problem#
Hydrometric data was available, but:
- No predictive analytics
- No structured data pipeline
- No automated model training
- Limited operational forecasting capability
Architecture#
- Data Storage: InfluxDB (15-second resolution river data)
- Monitoring: Grafana dashboards
- Preprocessing: Python data alignment & feature engineering
- Modeling: TensorFlow LSTM recurrent neural network
- Automation: GitLab CI/CD pipelines
- Deployment: Containerized services on Proxmox VM infrastructure
Key Engineering Elements#
- Multi-source time-series alignment (5-minute normalized intervals)
- Cyclical feature encoding for seasonality
- Automated model evaluation and metrics storage
- Forecast tagging for separation of observed vs predicted data
- Pipeline artifact retention and reporting
Results#
- 24-hour forecast generation
- Continuous retraining via CI pipeline
- Structured observability of prediction accuracy
- Production-ready telemetry-to-forecast workflow
Skills Demonstrated#
- Systems Architecture
- Infrastructure Automation
- Time-Series Data Engineering
- Machine Learning Integration
- Observability & Monitoring