Skip to content

Portfolio Architecture¶

3 ML projects deployed end-to-end — trained, containerized, served on GKE + EKS, monitored via Prometheus + Grafana.

YouTube Demo

System Overview¶

graph TB
    subgraph "CI/CD — GitHub Actions"
        GH[GitHub Actions] --> BUILD[Docker Build]
        BUILD --> AR[GCP Artifact Registry]
        BUILD --> ECR[AWS ECR]
    end

    subgraph "GCP — GKE (us-central1, 1-5 nodes auto-scaling)"
        GCE[nginx Ingress LB] --> BC1[BankChurn] & NL1[NLPInsight] & CT1[ChicagoTaxi]
        P1[Prometheus] --> G1[Grafana]
        D1[Drift CronJob] --> BC1
        M1[MLflow]
    end

    subgraph "AWS — EKS (us-east-1, 1-5 nodes auto-scaling)"
        AWS[nginx Ingress Classic ELB] --> BC2[BankChurn] & NL2[NLPInsight] & CT2[ChicagoTaxi]
        P2[Prometheus] --> G2[Grafana]
        D2[Drift CronJob] --> BC2
        M2[MLflow]
    end

    subgraph "IaC"
        TF[Terraform GCP+AWS] --> GCE & AWS
        KS[Kustomize Overlays] --> GCE & AWS
    end
Principle Implementation
Modularity Each project is self-contained with its own pipeline
Consistency Shared patterns: src/ layout, Pydantic config, FastAPI
Observability MLflow + Prometheus + Grafana
Security Gitleaks, Bandit, Trivy, pip-audit
Scalability Kubernetes HPA, CPU-based autoscaling
Multi-Cloud GKE + EKS with Kustomize overlays (ADR-013)

Projects (v3.5.3, Python 3.11.14 + sklearn 1.8.0)¶

Project Algorithm Primary Metric In-Pod Latency Tests Coverage
BankChurn StackingClassifier (RF+GB+XGB+LGB→LR) AUC 0.87 103ms p50 199 90%
NLPInsight TF-IDF + LogReg (prod) / FinBERT (GPU) Acc 80.6% 5ms p50 74 98%
ChicagoTaxi PySpark ETL + RandomForest (lag features) R² 0.96 75ms p50 122 91%

Infrastructure¶

  • Containers: Multi-stage Docker builds, non-root execution, --no-deps for heavy packages
  • Orchestration: GKE (GCP) + EKS (AWS), both live — Terraform IaC + Kustomize overlays
  • Model Delivery: GCS/S3 → Init Container → Pod (ConfigMap-driven paths)
  • CI/CD: GitHub Actions (deploy-gcp.yml + deploy-aws.yml) → build → push → deploy
  • Monitoring: Prometheus + Grafana + MLflow — cloud-agnostic, deployed on both clouds
  • Drift Detection: Daily CronJob checking health + prediction stability (PSI) across all services

GCP Production Cost (~$51/month)¶

Service Cost %
Compute Engine (4× e2-medium) $20.50 40%
GKE management $13.35 26%
Container Scanning $9.10 18%
Networking $6.15 12%
Cloud SQL (MLflow) $1.70 3%

Covered by Free Tier credits. Optimized with single-zone cluster, e2-medium nodes, and cleanup policies.

Technology Stack¶

Layer Technologies
ML scikit-learn 1.8.0, LightGBM 4.6+, HuggingFace Transformers, SHAP 0.50.0
API FastAPI, uvicorn, Pydantic
Tracking MLflow 3.10, DVC
Monitoring Prometheus, Grafana, Evidently AI, OpenTelemetry
Responsible AI Fairness audits (×3), drift detection (KS+PSI+Evidently), Pandera validation
Containers Docker, Kubernetes (GKE/EKS)
IaC Terraform (GCP + AWS modules)
CI/CD GitHub Actions, Trivy, Bandit, Gitleaks
Testing pytest (90–98% coverage, 395+ tests), Locust (load testing, 2,675 requests, 0% errors)

Visual Evidence¶

Multi-Cloud Deployment¶

GKE vs EKS (HERO) EKS Cluster Active EKS Workloads Running
Multi-Cloud EKS EKS Pods

GCP Production¶

GKE Workloads Grafana Monitoring MLflow Experiments
GKE Grafana MLflow

Live Demos¶

Demo File Description
Demo 01-demo-prediccion.gif ML predictions: BankChurn (SHAP) + NLPInsight + ChicagoTaxi
HPA 02-hpa-autoscaling.gif HPA auto-scaling under load (1→3 replicas)
Fairness 03-fairness-audit.gif Fairness audit CLI (disparate impact ratios)

Video: Portfolio Demo (3:30 min) — full multi-cloud walkthrough

Last Updated: March 2026 — Portfolio v3.5.3 (GKE + EKS deployed, NLB)