NASA Landslide Predictive Analysis
Time-Series Modeling with Automated MLOps
Problem
Geospatial risk models rot without fresh data, but re-training and deployment pipelines are often slow, fragile, and environment-inconsistent. The bottleneck isn't the model — it's the path from notebook to running service.
Approach
Provisioned the full training and serving environment in Terraform on AWS, so the infrastructure is one PR away from being torn down and rebuilt identically.
Containerized training and inference, then wired GitHub Actions and Jenkins for CI/CD so a model change goes through the same review-and-deploy path as application code.
Added basic observability — metrics, alerts, lineage — so a regression in production triggers a page, not a slack screenshot three days later.
Results
70% faster iteration · 100% IaC coverage
- ~70% faster iteration cycle (notebook → deployed model)
- 100% infrastructure-as-code coverage
- CI/CD parity between application and ML deploys
- Repeatable teardown / rebuild from a single Terraform plan
Stack
What I learned
On time-series risk problems, half the modeling battle is freshness. A slightly worse model retrained weekly beats a better model retrained quarterly. Optimizing the deploy loop is doing science, not avoiding it.