Case study · Production MLOps
Fitness MLOps Platform
A live, production-shaped stack: inference on Kubernetes, model lifecycle with MLflow, RAG with persisted vectors, autoscaling under load, and dashboards you can read like a senior review, not a docker-compose demo.
🚀 Production-ready📊 Grafana live☸️ EKS + HPA🔄 MLflow registry
Architecture
Vercel ▲ Next.js 15 → EKS ⚡ FastAPI → 📈 XGBoost + Chroma RAG → 📊 Grafana + Prometheus
▲
Next.js 15 (Vercel)
⚡
FastAPI (EKS)
📈
XGBoost + RAG
📊
Grafana / Prometheus
Production signals
HPA: 3 → 20 pods under load
MLflow model versions + registry workflow
Lambda cron retraining pipeline
StatefulSets for ChromaDB persistence
Dashboards: CPU, memory, predictions/sec
Repo layout (concept)
fitness-mlops/ ├── frontend/ # Next.js 15 ├── backend/ # FastAPI + XGBoost ├── k8s/ # Deployments, HPA, Services ├── mlflow/ # Model registry ├── lambda/ # Retraining cron ├── monitoring/ # Grafana + Prometheus └── .github/workflows/ # CI → ECR
Grafana-style targets
☸️
Pods (HPA)
3 active → 20 max
🖥️
CPU
~60% avg · 90% peak
⚡
Predictions/sec
15-45
🔄
MLflow models
3 versions
✅
Uptime
99.8%