Skip to content

ML/MLOps Portfolio — Production-Ready

3 ML Models • Multi-Cloud (GKE + EKS) • CI/CD • Prometheus + Grafana + MLflow

CI codecov Python Kubernetes Terraform

GitHub YouTube

Multi-Cloud: GKE vs EKS — Live

Multi-Cloud HERO: GKE vs EKS Side-by-Side

Same 6 services running on GCP (GKE, us-central1) and AWS (EKS, us-east-1) — multi-cloud deployed and verified.

ML Projects

Project Algorithm Metric Coverage Latency p50 (GCP / AWS)
BankChurn StackingClassifier (RF+GB+XGB+LGB→LR) + SHAP AUC 0.87 90% 200ms / 110ms
NLPInsight FinBERT Transformer (CPU) Acc 80.6% 98% 78ms / 100ms
ChicagoTaxi PySpark ETL (6.3M rows) + LightGBM R² 0.96 91% 100ms / 120ms

395+ tests across all projects, 0 failures, 85% CI threshold enforced.

Demo GIF

End-to-end: API predictions with SHAP explainability, sentiment analysis, and demand forecasting.

Production Infrastructure

Component GCP (GKE) AWS (EKS)
Cluster GKE 1 node baseline, auto-scales to 5 (e2-medium, us-central1) EKS 1 node baseline, auto-scales to 5 (t3.small, us-east-1)
Pods 6 Running, 0 restarts 6 Running, 0 restarts
Ingress nginx + static IP (136.111.152.72) nginx + NLB (AWS Load Balancer Controller)
Registry Artifact Registry ECR (3 repos)
Storage GCS (versioned) S3 (encrypted, versioned)
IAM Workload Identity IRSA
Drift Detection CronJob (daily 06:00 UTC) CronJob (daily 06:00 UTC)
Load Test 0% errors, p95 190ms 0% errors, p95 450ms
GKE Workloads (6 pods) EKS Cluster Active
GKE EKS

Observability Stack

Grafana ML Dashboard Prometheus Targets (UP) MLflow Experiments
Grafana Prometheus MLflow
  • Prometheus: 16 targets UP, 16 alert rules, 15s scrape interval
  • Grafana: 2 dashboards, 26 panels (latency, throughput, predictions, errors, resources)
  • MLflow: 9 experiments tracked across 3 projects

CI/CD Pipeline

GitHub Actions (10 jobs) Deploy Workflows (GCP + AWS) Codecov Dashboard
Pipeline Deploy Codecov
  • CI: tests (matrix) → security (Gitleaks, Bandit) → Docker build (Trivy) → integration tests
  • CD: deploy-gcp.yml + deploy-aws.yml — automated multi-cloud deployment
  • Coverage: 90–98% enforced at 85% threshold via Codecov

Infrastructure as Code

Terraform Multi-Cloud K8s Overlays (GCP vs AWS) Infra Tests
Terraform Overlays Tests
  • Terraform: GCP + AWS modules, remote state, terraform plan = no drift
  • Kustomize: Shared base manifests + cloud-specific overlays
  • Testing: tfsec + checkov + conftest (OPA) + kube-linter — 9/9 GCP, 8/8 AWS

API Evidence

BankChurn (SHAP) NLPInsight (Sentiment) ChicagoTaxi (Demand)
BankChurn NLPInsight ChicagoTaxi

All services expose FastAPI with Swagger UI, Prometheus /metrics, and structured health checks.

Security & Automation

Feature Status
Pod Security Standards enforce=baseline, warn=restricted
Network Policies default-deny + 3 allow rules
Pod Disruption Budgets minAvailable=1 (3 ML services)
Drift Detection Daily CronJob on both clouds
Retraining Trigger Automated CronJob → GitHub Actions dispatch
Scanning Gitleaks + Bandit + Trivy + pip-audit (blocking on HIGH)
Non-root All containers run as UID 1000

HPA Auto-Scaling

HPA Auto-Scaling

CPU-based HPA: 1→3 replicas under load, automatic scale-down after traffic subsides.

Service CPU Target Pods Memory
BankChurn 70% 1–3 ~396Mi
NLPInsight 75% 1–3 ~283Mi
ChicagoTaxi 70% 1–3 ~288Mi

Responsible AI

Fairness Audit

  • Fairness Audits: Disparate impact ratio + equal opportunity (BankChurn by Gender/Geography)
  • Drift Detection: KS + PSI + Evidently per feature, automated alerting
  • Data Validation: Pandera schemas for all projects (training + inference)
  • Adversarial Testing: 43 robustness tests (SQL injection, XSS, boundary, Unicode)

Technology Stack

Layer Technologies
ML/DS scikit-learn 1.8.0, XGBoost, LightGBM, PySpark, Dask, SHAP, Optuna
API FastAPI, Pydantic, uvicorn (2 workers)
MLOps MLflow, DVC, Evidently AI, OpenTelemetry
Cloud GCP (GKE, GCS, AR, Cloud SQL), AWS (EKS, S3, ECR)
IaC Terraform (GCP + AWS), Kustomize overlays
Monitoring Prometheus, Grafana (25 panels), 16 alert rules
CI/CD GitHub Actions (CI + deploy-gcp + deploy-aws), Codecov
Security Gitleaks, Bandit, Trivy, pip-audit, Network Policies, PDBs
Testing pytest (295+ tests, 90–98%), Locust load testing

Quick Start

git clone https://github.com/DuqueOM/ML-MLOps-Portfolio.git
cd ML-MLOps-Portfolio
bash scripts/setup_demo_models.sh
docker compose -f docker-compose.demo.yml up -d --build
Service URL
BankChurn API localhost:8001/docs
NLPInsight API localhost:8003/docs
ChicagoTaxi API localhost:8004/docs
MLflow UI localhost:5000
**Built by [Duque Ortega Mutis](https://github.com/DuqueOM)** | [LinkedIn](https://linkedin.com/in/duqueom) | [Video Demo](https://youtu.be/7dFFqq2ROPw) *Portfolio v3.5.3 — March 2026 — Deployed on GCP (GKE) + AWS (EKS)*