Projects Overview¶

Three ML systems built end-to-end: trained, containerized, deployed on Kubernetes, and monitored in production.

Portfolio Demo

System Architecture¶

graph LR
    subgraph "Kubernetes Cluster (GKE / EKS)"
        ING[nginx Ingress] --> BC[BankChurn :8001]
        ING --> NLP[NLPInsight :8003]
        ING --> CT[ChicagoTaxi :8004]
        BC --> PROM[Prometheus]
        NLP --> PROM
        CT --> PROM
        PROM --> GRAF[Grafana]
    end

Comparison¶

Aspect	BankChurn	NLPInsight	ChicagoTaxi
Domain	Banking — customer churn	Finance — sentiment	Urban mobility — demand
Algorithm	StackingClassifier (RF+GB+XGB+LGB→LR)	TF-IDF + LogReg / FinBERT	PySpark ETL + RandomForest
Primary Metric	AUC 0.87	Accuracy 80.6%	R² 0.9649
Why This Metric	Imbalanced (20% churn): AUC ranks correctly	3-class finance text: Accuracy + F1 for per-class balance	Continuous demand: R² captures variance explained
Latency (p50)	200ms (GCP) / 110ms (AWS)	78ms (GCP) / 100ms (AWS)	100ms (GCP) / 120ms (AWS)
Docker Image	490 MB	2.1 GB	382 MB
Tests / Coverage	199 / 90%	74 / 98%	22 / 91%
Key Feature	SHAP explainability	Dual-backend auto-detection	6.3M row PySpark pipeline

Live Evidence — Multi-Cloud¶

GKE Workloads (GCP)	EKS Workloads (AWS)

kubectl Pods (GCP)	kubectl Pods (AWS)	Resource Usage

API Predictions — Live¶

BankChurn SHAP	NLPInsight Sentiment	ChicagoTaxi Demand

Architecture Decisions¶

Each project makes deliberate trade-offs documented in ADRs:

BankChurn: StackingClassifier over simpler models for +5 AUC points; KernelExplainer for SHAP (4.5s latency accepted for explainability)
NLPInsight: TF-IDF for production (5ms), FinBERT available for GPU environments
ChicagoTaxi: Leak-free lag features with temporal split validation; Dask for batch serving

Links¶

BankChurn Predictor — threshold tuning, cost analysis, SHAP explainability
NLPInsight Analyzer — TF-IDF production / FinBERT GPU, dual backend
ChicagoTaxi Demand Pipeline — PySpark ETL, Dask batch prediction, 6.3M rows

Last Updated: March 2026 — v3.5.3