All case studies
2025
MLOps Inference Pipeline with Monitoring
Production-grade MLOps pipeline with automated model training, a FastAPI inference API, and a monitoring stack of Prometheus, Grafana and Alertmanager wired to Slack, provisioned on AWS with Terraform.
Stack
- FastAPI
- Prometheus
- Grafana
- Docker
- AWS
- Terraform
- Python
- MLflow
- DVC
- MLOps
Problem
A model in production needs observability, alerting, and a route back to retraining when the data moves underneath it. Without those it is not deployed, it is merely running.
Architecture
Terraform -> EC2. Docker services: FastAPI inference, Prometheus, Grafana, Alertmanager -> Slack.
What it does
- Prometheus and Grafana monitoring with Slack alerting
- AWS EC2 provisioned through Terraform as infrastructure as code
- Drift detection triggering automated retraining
- Multi-service Docker architecture
Challenges & learnings
- Choosing ML metrics that mean something in Prometheus: latency, throughput and error rate rather than accuracy, which does not move at inference time.
- Terraform state handling and safe EC2 lifecycle management.