Operational ML Solutions
FrostPulse MetarPulse Contact

Lightweight
Edge-Ready
Machine Learning
Operational
Production

Production-proven ML systems for atmospheric forecasting. Edge-deployed on commodity hardware with measurable accuracy metrics across agriculture and aviation safety applications.

FrostPulse
 •
MetarPulse
🌡️ Agricultural Intelligence
📄 Pitch Estudio Técnico (ES)

FrostPulse

Operational frost prediction system for Chilean agriculture achieving 97-99% accuracy across validated meteorological stations. Edge-deployed on commodity hardware for real-time farm protection with 6-12 hour forecast windows.

97-99% Prediction Accuracy
~900KB Model Size
6-12h Forecast Window
Edge Deployment
✈️ Aviation Safety Suite
📄 View Visibility Research on arXiv →

MetarPulse

Comprehensive aviation safety suite powered by standard METAR observations. Combines physics-informed visibility nowcasting with low-level wind shear detection—lightweight alternatives to expensive radar infrastructure, validated across multiple continents and climatic regimes.

IFR Visibility Nowcasting

Physics-informed aviation visibility forecasting achieving 0.89-0.98 AUC across 11 international airports spanning multiple climatic regimes, with 2.5-4x improvement over operational forecasts.

0.89-0.98 AUC Range
2.5-4x Improvement
11 Airports
~900KB Model Size
  • Geographic Transferability: Zero-shot deployment across tropical, temperate, Mediterranean, and continental climates
  • Coordinate-Free Architecture: 14 physics-informed features capture universal atmospheric processes (radiative cooling, dewpoint depression, thermal gradients)
  • Operational Impact: 2-hour lead time tactical forecasting for flight scheduling and ground operations

Low-Level Wind Shear Detection

METAR-based wind shear detection validated across 12 international airports, 6 continents, 3.4M observations spanning 24 years. Lightweight alternative to Terminal Doppler Weather Radar ($5-10M infrastructure).

0.75-0.89 AUC Range
>75% Recall (Safety Mode)
12 Airports
5-min METAR Compatible
  • TDWR Alternative: Surface-based detection using standard METAR vs $5-10M radar infrastructure
  • High-Frequency Operations: Validated with 5-minute automated METAR (0.881 AUC), supporting real-time nowcasting
  • Global Validation: Tested across polar to desert climates, demonstrating geographic transferability

Integrated Deployment Advantages

  • Single Data Source: Both capabilities use standard METAR observations—no additional sensors required
  • Unified Edge Platform: Deploy both systems on commodity airport hardware with shared data pipeline
  • SHAP Interpretability: Models learn transferable atmospheric physics, providing operational transparency
  • Commodity Hardware: Runs on standard x86 servers, no GPU or specialized infrastructure required
  • Operational Accessibility: Enables advanced aviation safety for resource-constrained airports (regional, island, polar stations)

Independent ML Research & Consulting

Marcelo Cerda Castillo is an independent researcher and ML engineering consultant with published research on atmospheric forecasting systems. Former IT Director at Chile's Meteorological Directorate (DMC, 2002-2012), specializing in operational ML deployment for agriculture and aviation safety.

Research published on arXiv with validation across multiple continents and climate regimes. Consulting services include custom atmospheric forecasting systems, physics-informed ML architectures, and edge-deployed operational solutions optimized for commodity hardware.

Discuss Your ML Challenge