E-Commerce Conversion Intelligence Platform

End-to-end machine learning, business intelligence, and MLOps for purchase-intent prediction

Transforming raw e-commerce behaviour into visitor-level intelligence, campaign-ready audiences, explainable model decisions, monitored production signals, and deployable business applications.

Launch the app · View CI/CD · Explore the source

Navigate: Overview · Architecture · Application · MLOps · Monitoring · CI/CD · Run Locally

Executive Summary

The E-Commerce Conversion Intelligence Platform is a portfolio-grade, end-to-end machine learning and MLOps system built to answer one practical business question:

Which visitors are most likely to purchase, and how should the business act on that signal?

The platform converts raw RetailRocket event data into visitor-level features, benchmarks multiple models, selects and validates a champion, optimizes the decision threshold, scores visitors, creates campaign-ready audiences, explains model decisions, monitors drift and delayed outcomes, and presents the complete workflow through a multi-page Streamlit business application.

This is not only a predictive model. It is a complete analytics product that connects:

business problem framing;
event-level data engineering;
rare-event classification;
campaign prioritization;
model explainability;
experiment tracking;
drift monitoring;
delayed-label performance measurement;
business forecasting;
anomaly detection;
customer segmentation and journey intelligence;
containerization and orchestration;
CI/CD, security, and reproducibility controls.

Project at a Glance

Metric	Result	Business meaning
Visitors analyzed	1,407,500	Full visitor population evaluated
Observed conversion rate	0.827%	Highly imbalanced purchase problem
High-intent audience	19,588 visitors	Focused campaign activation group
High-intent share	~1.39%	Small, prioritized audience
Champion model	Tuned Random Forest	Final deployable purchase-intent model
PR-AUC	0.4151	Strong ranking quality for rare conversions
ROC-AUC	0.9696	Strong class separation
Precision	36.61%	Selected audience is much cleaner than random targeting
Recall	72.16%	Captures a large share of actual buyers
F1 score	0.4858	Balanced precision and recall
Decision threshold	0.97	Business cutoff for high-intent selection
Anomaly rate	3.66%	Share of visitors flagged for unusual behaviour
Buyer concentration	~44× baseline	Model precision versus observed conversion rate

Important: the ~44× concentration is model-evaluation evidence, not proven campaign uplift. Real marketing impact must still be validated with controlled A/B testing.

Why This Project Matters

Most e-commerce visitors do not purchase. Broad targeting wastes budget on users with little evidence of intent.

This platform changes the decision process from:

Traditional workflow	Intelligence-driven workflow
Target a broad visitor population	Prioritize high-intent visitors
Use traffic volume as the main signal	Use visitor-level purchase probability
Select campaigns manually	Use model-supported audience ranking
Monitor only application uptime	Monitor features, predictions, outcomes, and services
Treat ML as a notebook result	Deliver ML through a governed business application

The final output is not simply a probability. It is a decision system:

Decision flow: raw behaviour becomes an actionable audience, then measured outcomes become new monitoring evidence.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    subgraph SIGNAL["1 · Build the signal"]
        direction LR
        A["Raw visitor<br/>events"]:::data --> B["Visitor-level<br/>features"]:::data
        B --> C["Purchase-intent<br/>model"]:::ml
    end

    subgraph DECIDE["2 · Make the decision"]
        direction LR
        D["Calibrated<br/>intent score"]:::ml --> E{"Business<br/>threshold"}:::decision
        E --> F["High-intent<br/>audience"]:::business
    end

    subgraph LEARN["3 · Measure and improve"]
        direction LR
        G["Campaign<br/>prioritization"]:::business --> H["Measured<br/>outcomes"]:::ops
        H --> I["Monitoring and<br/>retraining evidence"]:::monitor
    end

    C --> D
    F --> G

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Business Capabilities

The platform supports six connected decision areas:

Capability	Business output
Purchase-intent prediction	Probability that a visitor will convert
Campaign prioritization	Ranked and capacity-aware high-intent audience
Model governance	Champion/challenger comparison, threshold, stability, and economics
Customer intelligence	Segments, nearest-neighbour context, and journey patterns
Operational intelligence	Forecasts, anomalies, drift, delayed outcomes, and system health
Deployment intelligence	Streamlit Cloud, Docker Compose, Kubernetes, Helm, and CI/CD evidence

End-to-End Platform Architecture

Six connected layers move the project from raw events to a governed business application.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    CI["GitHub Actions<br/>quality · security · delivery"]:::platform

    subgraph DATA["1 · Data foundation"]
        direction LR
        D1["RetailRocket<br/>events"]:::data --> D2["Clean and<br/>validate"]:::data
        D2 --> D3["Visitor-level<br/>feature table"]:::data
    end

    subgraph ML["2 · Machine-learning decision engine"]
        direction LR
        M1["Benchmark<br/>candidates"]:::ml --> M2["Tuned Random<br/>Forest"]:::ml
        M2 --> M3["Threshold and<br/>batch scores"]:::decision
    end

    subgraph INTEL["3 · Business intelligence"]
        direction LR
        B1["Executive and<br/>visitor intelligence"]:::business
        B2["Campaign and<br/>model intelligence"]:::business
        B3["Forecast, anomaly<br/>and journey"]:::business
    end

    subgraph GOVERN["4 · MLOps and governance"]
        direction LR
        O1["MLflow<br/>experiments"]:::ops
        O2["Model metadata<br/>and provenance"]:::ops
        O3["Prediction ledger<br/>and delayed labels"]:::ops
        O4["Evidently<br/>drift analysis"]:::ops
    end

    subgraph OBS["5 · Observability"]
        direction LR
        P1["Monitoring<br/>snapshot"]:::monitor --> P2["Prometheus and<br/>health probes"]:::monitor
        P2 --> P3["Grafana and<br/>Alertmanager"]:::monitor
    end

    subgraph DELIVERY["6 · Delivery"]
        direction LR
        X1["Streamlit<br/>Community Cloud"]:::platform
        X2["Docker<br/>Compose"]:::platform
        X3["Kubernetes<br/>and Helm"]:::platform
    end

    CI -. validates .-> DATA
    D3 --> M1
    M3 --> B1
    M3 --> B2
    M3 --> B3
    M2 --> O1
    M3 --> O2
    M3 --> O3
    D3 --> O4
    O3 --> P1
    O4 --> P1
    B1 --> X1
    P3 --> X2
    X2 --> X3

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Data Engineering Pipeline

The RetailRocket dataset is event-level: one row represents one visitor action such as a product view, add-to-cart event, or transaction. The business decision is visitor-level, so the project changes the data grain.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    subgraph INGEST["1 · Ingest and verify"]
        direction LR
        A["events.csv<br/>one action per row"]:::data --> B["Schema and<br/>quality checks"]:::data
        B --> C["Clean event<br/>records"]:::data
    end

    subgraph AGG["2 · Change the grain"]
        direction LR
        D["Group by<br/>visitor"]:::ml --> E["Create behavioural<br/>features"]:::ml
        E --> F["One row<br/>per visitor"]:::ml
    end

    subgraph READY["3 · Prepare for modelling"]
        direction LR
        G["Training and<br/>scoring split"]:::decision --> H["Model-ready<br/>dataset"]:::business
    end

    C --> D
    F --> G

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Core visitor features

Feature	Meaning
`view_count`	Number of product views
`addtocart_count`	Number of add-to-cart actions
`unique_items`	Number of distinct items viewed
`activity_span_ms`	Duration of observed visitor activity
`converted`	Purchase outcome used as the target

The project also contains modular packages for:

feature engineering;
segmentation;
forecasting;
anomaly detection;
model evaluation;
monitoring and delayed labels;
visualization and business intelligence.

Machine Learning Strategy

The observed conversion rate is only 0.827%, so accuracy is not a useful primary metric. A model that predicts “no purchase” for almost everyone can look accurate while being useless for campaign targeting.

The project therefore emphasizes:

Metric	Why it matters
PR-AUC	Evaluates ranking quality under severe class imbalance
ROC-AUC	Measures buyer versus non-buyer separation
Precision	Measures the quality of the selected campaign audience
Recall	Measures how many actual buyers are captured
F1	Balances precision and recall
Threshold economics	Converts model scores into a business operating decision

Model lifecycle

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    subgraph DEVELOP["1 · Develop"]
        direction LR
        A["Visitor feature<br/>table"]:::data --> B["Baseline<br/>model"]:::ml
        A --> C["Candidate<br/>benchmark"]:::ml
        C --> D["Hyperparameter<br/>tuning"]:::ml
    end

    subgraph SELECT["2 · Select and govern"]
        direction LR
        E["Tuned Random<br/>Forest champion"]:::ml --> F["Threshold<br/>analysis"]:::decision
        F --> G["Stability and<br/>sensitivity checks"]:::decision
        G --> H["Champion<br/>metadata"]:::ops
    end

    subgraph OPERATE["3 · Score and learn"]
        direction LR
        I["Batch visitor<br/>scoring"]:::business --> J["Campaign<br/>audience"]:::business
        J --> K["Prediction<br/>ledger"]:::ops
        K --> L["Delayed outcome<br/>evaluation"]:::monitor
    end

    D --> E
    H --> I

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Champion decision

The tuned Random Forest was selected because it combined strong rare-event ranking with a practical campaign operating point:

PR-AUC: 0.4151
ROC-AUC: 0.9696
Precision: 36.61%
Recall: 72.16%
F1: 0.4858
Decision threshold: 0.97

Streamlit Business Application

The Streamlit application is the business-facing layer of the platform.

Entrypoint: app/Executive_Overview.py

Navigation model: one executive command centre connects decision, understanding, and governance workflows.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    HOME["Executive Overview<br/>portfolio command centre"]:::business

    subgraph DECIDE["Decide"]
        direction LR
        P1["Visitor Intent Predictor<br/>single-visitor support"]:::business
        P2["Batch Scoring<br/>audience and capacity"]:::business
        P3["Model Benchmark Selection<br/>champion and economics"]:::business
    end

    subgraph UNDERSTAND["Understand"]
        direction LR
        P4["Business KPI Forecasting<br/>scenarios and residuals"]:::ml
        P5["Anomaly and Outlier<br/>triage and investigation"]:::ml
        P9["Customer Segmentation<br/>segments and journeys"]:::ml
    end

    subgraph GOVERN["Govern"]
        direction LR
        P6["Monitoring, Drift and Health<br/>data, outcomes and services"]:::monitor
        P7["MLOps Architecture<br/>components and deployment"]:::ops
        P8["ML Validation and Evidence<br/>tests and provenance"]:::ops
    end

    HOME --> DECIDE
    HOME --> UNDERSTAND
    HOME --> GOVERN

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Application pages

Page	Primary purpose
Executive Overview	Headline KPIs, business narrative, risk, and action priorities
Visitor Intent Predictor	Score and explain a single visitor scenario
Batch Scoring	Score visitors, apply campaign capacity, and create target audiences
Model Benchmark Selection	Compare candidates and explain the champion decision
Business KPI Forecasting	Forecast business KPIs with scenario and residual diagnostics
Anomaly and Outlier	Detect, rank, and investigate unusual behaviour
Monitoring, Drift and Health	Review data drift, prediction drift, delayed outcomes, and service status
MLOps Architecture	Explain platform components, deployment paths, and operational boundaries
ML Validation and Evidence	Surface test, reproducibility, provenance, and model-validation evidence
Customer Segmentation and Journey	Explore behavioural segments, similar visitors, and journey patterns

Campaign Intelligence

Batch scoring is designed as a business workflow, not only a technical inference step.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    subgraph INPUT["1 · Prepare the audience"]
        direction LR
        A["Visitor<br/>batch"]:::data --> B["Feature<br/>validation"]:::data
        B --> C["Champion<br/>scoring"]:::ml
    end

    subgraph PRIORITIZE["2 · Prioritize"]
        direction LR
        D["Intent<br/>probability"]:::ml --> E{"Threshold<br/>decision"}:::decision
        E --> F["Capacity-aware<br/>ranking"]:::decision
    end

    subgraph ACTIVATE["3 · Activate"]
        direction LR
        G["Campaign<br/>audience"]:::business --> H["Recommended<br/>action"]:::business
        H --> I["Export and<br/>activation"]:::ops
    end

    C --> D
    F --> G

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

The app supports:

score distributions;
campaign capacity scenarios;
audience-size trade-offs;
threshold effects;
high-intent segment composition;
conversion-risk context;
action recommendations;
exportable campaign lists.

Explainability and Similarity Intelligence

The platform combines global model evidence with local visitor-level context.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    A["Scored visitor"]:::business

    subgraph MODEL["Model evidence"]
        direction LR
        B["Feature<br/>values"]:::data --> F["Feature contribution<br/>explanation"]:::ml
        C["Nearest real<br/>visitors"]:::data --> G["Similarity and<br/>difference breakdown"]:::ml
    end

    subgraph CONTEXT["Business context"]
        direction LR
        D["Segment<br/>membership"]:::business --> H["Segment<br/>profile"]:::business
        E["Anomaly<br/>context"]:::monitor --> I["Risk<br/>flags"]:::monitor
    end

    J["Governed visitor explanation<br/>score · evidence · context · caution"]:::decision

    A --> B
    A --> C
    A --> D
    A --> E
    F --> J
    G --> J
    H --> J
    I --> J

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

This supports more responsible business use by showing:

why a visitor received a score;
which behaviours are most influential;
how the visitor compares with similar real visitors;
whether the visitor belongs to a specific behavioural segment;
whether anomaly signals require additional review.

MLflow Experiment Tracking Architecture

MLflow is isolated from the lightweight app runtime and uses its own pinned environment.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    A["Training and<br/>benchmark scripts"]:::data --> B["MLflow bridge"]:::ml

    subgraph LOG["1 · Log the experiment"]
        direction LR
        C["Run<br/>parameters"]:::ml
        D["Evaluation<br/>metrics"]:::ml
        E["Model<br/>artifacts"]:::ml
        F["Environment<br/>provenance"]:::ml
    end

    subgraph STORE["2 · Store the evidence"]
        direction LR
        G["MLflow tracking<br/>server"]:::ops --> H["SQLite tracking<br/>backend"]:::ops
        G --> I["Artifact<br/>store"]:::ops
    end

    J["Experiment<br/>comparison"]:::decision --> K["Champion decision<br/>evidence"]:::business

    B --> C
    B --> D
    B --> E
    B --> F
    C --> G
    D --> G
    E --> G
    F --> G
    H --> J
    I --> J

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

MLflow responsibilities

record model parameters and evaluation metrics;
preserve experiment and environment provenance;
validate model serialization and loading;
support champion/challenger evidence;
provide an auditable bridge between training and deployment metadata.

Isolated dependency file: requirements-mlflow.txt

Validated MLflow version: 3.14.0

Evidently Drift and Monitoring Architecture

Evidently is used in an isolated monitoring environment to avoid dependency conflicts with the application runtime.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    subgraph INPUTS["1 · Compare two populations"]
        direction LR
        R["Reference<br/>features + predictions"]:::data
        C["Current<br/>features + predictions"]:::data
    end

    E["Evidently<br/>analysis bridge"]:::ml

    subgraph REPORTS["2 · Produce drift evidence"]
        direction LR
        F["Feature drift<br/>report"]:::monitor
        G["Prediction drift<br/>report"]:::monitor
        H["Drift summary<br/>and provenance"]:::monitor
    end

    subgraph ACTIONS["3 · Publish and act"]
        direction LR
        I["Streamlit<br/>monitoring page"]:::business
        J["Monitoring<br/>snapshot"]:::ops
        K["Prometheus, Grafana<br/>and alerts"]:::monitor
    end

    R --> E
    C --> E
    E --> F
    E --> G
    F --> H
    G --> H
    H --> I
    H --> J
    J --> K

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Evidently responsibilities

compare reference and current feature distributions;
evaluate prediction drift;
summarize drifted features and affected populations;
preserve report provenance;
feed drift evidence into the monitoring layer and Streamlit application.

Isolated dependency file: requirements-evidently.txt

Validated Evidently version: 0.7.21

Delayed Labels and Production Performance

Purchase outcomes may arrive after predictions are made. The project therefore separates prediction-time evidence from later outcome evaluation.

%%{init: {"theme":"base","sequence":{"useMaxWidth":true,"wrap":true,"diagramMarginX":30,"diagramMarginY":20,"actorMargin":45,"messageMargin":35},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","actorBkg":"#E0F2FE","actorBorder":"#0284C7","actorTextColor":"#0F172A","signalColor":"#475569","signalTextColor":"#0F172A","noteBkgColor":"#FEF3C7","noteBorderColor":"#D97706","noteTextColor":"#78350F"}}}%%
sequenceDiagram
    autonumber
    participant App as Scoring App
    participant Ledger as Prediction Ledger
    participant Outcome as Delayed Outcome
    participant Eval as Production Evaluation
    participant Monitor as Monitoring

    App->>Ledger: Store visitor, score, threshold and model version
    Note over Ledger: Prediction-time evidence is preserved
    Outcome->>Ledger: Attach the later purchase outcome
    Ledger->>Eval: Build the matched evaluation sample
    Eval->>Eval: Calculate precision, recall and conversion evidence
    Eval->>Monitor: Publish production-performance summary
    Monitor-->>App: Display health and performance evidence

This prevents a common MLOps mistake: evaluating only training metrics while ignoring real post-deployment outcomes.

Monitoring and Observability Stack

The full local monitoring stack includes application, model, and infrastructure signals.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    subgraph SIGNALS["1 · Operational signals"]
        direction LR
        APP["Streamlit<br/>application"]:::business
        LEDGER["Prediction and<br/>delayed-label outputs"]:::data
        DRIFT["Evidently drift<br/>summaries"]:::data
    end

    subgraph COLLECT["2 · Collect and expose"]
        direction LR
        BLACKBOX["Blackbox Exporter<br/>:9115"]:::ops
        SNAPSHOT["Monitoring<br/>snapshot"]:::ops
        EXPORTER["Metrics Exporter<br/>:8000"]:::ops
    end

    PROM["Prometheus<br/>:9090"]:::monitor

    subgraph ACT["3 · Visualize and respond"]
        direction LR
        GRAFANA["Grafana<br/>:3000"]:::monitor
        ALERT["Alertmanager<br/>:9093"]:::monitor
        WEBHOOK["Alert Webhook<br/>:5001"]:::monitor
    end

    APP --> BLACKBOX
    LEDGER --> SNAPSHOT
    DRIFT --> SNAPSHOT
    SNAPSHOT --> EXPORTER
    BLACKBOX --> PROM
    EXPORTER --> PROM
    PROM --> GRAFANA
    PROM --> ALERT
    ALERT --> WEBHOOK

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Monitored areas

Area	Examples
Application health	Streamlit availability and probe status
Model output	Score distribution, high-intent rate, threshold outcomes
Data quality	Missingness, invalid values, population changes
Drift	Feature drift and prediction drift
Production outcomes	Delayed-label precision, recall, and conversion evidence
Business health	Campaign audience size, forecast movement, anomaly rate
Infrastructure	Metrics endpoint health, alert routing, dashboard availability

Monitoring snapshot pattern

Prometheus should scrape lightweight metrics, not repeatedly scan large analytical files. The project therefore uses a cached snapshot:

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    A["Large analytical<br/>CSV and JSON outputs"]:::data --> B["Snapshot<br/>builder"]:::ml
    B --> C["ecommerce_metrics_<br/>snapshot.json"]:::ops
    C --> D["Lightweight metrics<br/>exporter"]:::ops
    D --> E["Prometheus"]:::monitor

    subgraph CONSUMERS["Fast monitoring consumers"]
        direction LR
        F["Grafana<br/>dashboards"]:::monitor
        G["Alertmanager<br/>rules"]:::monitor
    end

    E --> F
    E --> G

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Docker Compose Architecture

The local platform runs as a multi-service Docker Compose stack.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    USER["Business user"]:::business --> APP["streamlit-app<br/>:8501"]:::business

    subgraph DATA["Read-only analytical assets"]
        direction LR
        D1["Data"]:::data
        D2["Reports"]:::data
        D3["Models"]:::data
    end

    subgraph METRICS["Monitoring services"]
        direction LR
        CACHE["Metrics<br/>cache"]:::ops --> EXPORTER["metrics-exporter<br/>:8000"]:::ops
        BLACKBOX["blackbox-exporter<br/>:9115"]:::ops
        PROM["Prometheus<br/>:9090"]:::monitor
    end

    subgraph RESPONSE["Dashboards and alerts"]
        direction LR
        GRAFANA["Grafana<br/>:3000"]:::monitor
        ALERTMANAGER["Alertmanager<br/>:9093"]:::monitor
        WEBHOOK["alert-webhook<br/>:5001"]:::monitor
    end

    D1 --> APP
    D2 --> APP
    D3 --> APP
    APP --> BLACKBOX
    BLACKBOX --> PROM
    EXPORTER --> PROM
    PROM --> GRAFANA
    PROM --> ALERTMANAGER
    ALERTMANAGER --> WEBHOOK

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Docker services

Service	Purpose	Local port
`streamlit-app`	Business application	`8501`
`metrics-exporter`	Prometheus-compatible project metrics	`8000`
`prometheus`	Metrics collection and rule evaluation	`9090`
`blackbox-exporter`	External application health probes	`9115`
`alertmanager`	Alert grouping and routing	`9093`
`alert-webhook`	Local alert receiver and evidence writer	`5001`
`grafana`	Mission-control dashboards	`3000`

CI/CD and Quality Gates

The GitHub Actions workflow validates code, security, machine learning integrations, containers, orchestration files, and monitoring configuration.

Pipeline design: quality and security gates run before ML validation and deployment-asset checks.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    A["Push · pull request<br/>manual · weekly"]:::platform --> B["GitHub Actions"]:::platform

    subgraph QUALITY["1 · Code quality and security"]
        direction LR
        Q1["Python 3.10<br/>dependencies"]:::data --> Q2["Imports, structure<br/>and syntax"]:::data
        Q2 --> Q3["Ruff lint<br/>and format"]:::data
        Q3 --> Q4["pip-audit and<br/>secret scan"]:::data
    end

    subgraph MLTEST["2 · Machine-learning validation"]
        direction LR
        M1["Sample-data<br/>smoke training"]:::ml --> M2["pytest<br/>suite"]:::ml
        M2 --> M3["MLflow<br/>integration"]:::ml
        M2 --> M4["Evidently<br/>integration"]:::ml
    end

    subgraph DELIVERY["3 · Delivery validation"]
        direction LR
        D1["Docker image and<br/>Compose"]:::ops --> D2["Helm and<br/>Kubernetes"]:::ops
        D2 --> D3["Prometheus and<br/>Alertmanager"]:::monitor
        D3 --> D4["Grafana JSON<br/>validation"]:::monitor
    end

    U["Green pipeline<br/>or compact failure evidence"]:::business

    B --> Q1
    Q4 --> M1
    M3 --> D1
    M4 --> D1
    D4 --> U

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

CI controls

Python 3.10 dependency installation and pip check;
project import and folder validation;
Python and shell syntax checks;
delayed-label and production-performance tests;
Ruff linting and format checks;
dependency vulnerability scanning with pip-audit;
repository secret scanning;
sample-data smoke training;
complete automated test suite;
scheduled/manual MLflow integration validation;
scheduled/manual Evidently integration validation;
Docker image and Docker Compose validation;
Helm linting and rendered Kubernetes validation;
Prometheus, alert-rule, Alertmanager, and Grafana validation;
compact failure-evidence uploads.

The CI workflow validates deployment assets but does not automatically deploy production infrastructure.

Deployment Paths

The repository supports three delivery modes.

%%{init: {"theme":"base","flowchart":{"curve":"basis","htmlLabels":true,"nodeSpacing":32,"rankSpacing":48},"themeVariables":{"fontFamily":"Inter,Segoe UI,Arial,sans-serif","fontSize":"16px","primaryColor":"#E0F2FE","primaryTextColor":"#0F172A","primaryBorderColor":"#0284C7","lineColor":"#64748B","secondaryColor":"#EDE9FE","tertiaryColor":"#DCFCE7","clusterBkg":"#F8FAFC","clusterBorder":"#94A3B8","edgeLabelBackground":"#FFFFFF"}}}%%
flowchart TB
    GIT["GitHub<br/>main branch"]:::platform --> CI["GitHub Actions<br/>validation"]:::platform

    subgraph PATHS["Three supported delivery paths"]
        direction LR

        subgraph CLOUDPATH["Portfolio cloud"]
            C1["Streamlit<br/>Community Cloud"]:::business --> C2["Public portfolio<br/>application"]:::business
        end

        subgraph LOCALPATH["Local MLOps stack"]
            L1["Docker<br/>image"]:::ops --> L2["Docker<br/>Compose"]:::ops
            L2 --> L3["App + monitoring<br/>services"]:::ops
        end

        subgraph K8SPATH["Orchestrated demo"]
            K1["Kubernetes<br/>manifests"]:::monitor --> K2["Helm<br/>chart"]:::monitor
            K2 --> K3["Kubernetes<br/>demo release"]:::monitor
        end
    end

    CI --> C1
    CI --> L1
    CI --> K1

classDef data fill:#E0F2FE,stroke:#0284C7,color:#0C4A6E,stroke-width:1.5px;
    classDef ml fill:#EDE9FE,stroke:#7C3AED,color:#4C1D95,stroke-width:1.5px;
    classDef business fill:#DCFCE7,stroke:#16A34A,color:#14532D,stroke-width:1.5px;
    classDef ops fill:#FFEDD5,stroke:#EA580C,color:#7C2D12,stroke-width:1.5px;
    classDef monitor fill:#FFE4E6,stroke:#E11D48,color:#881337,stroke-width:1.5px;
    classDef platform fill:#F1F5F9,stroke:#64748B,color:#1E293B,stroke-width:1.5px;
    classDef decision fill:#FEF3C7,stroke:#D97706,color:#78350F,stroke-width:1.5px;
    linkStyle default stroke:#64748B,stroke-width:1.5px;

Streamlit Community Cloud

Repository: RuturajM31/E-Commerce-Conversion-Intelligence-Platform
Branch: main
Entrypoint: app/Executive_Overview.py
Validated Python: 3.10
App dependencies: app/requirements.txt
Secrets required: No
Public URL: https://ecommerce-conversion-intelligence.streamlit.app

Docker Compose

Build the monitoring snapshot before starting the stack:

python3 -m src.monitoring.build_monitoring_snapshot
docker compose up -d --build
docker compose ps

Open:

Streamlit: http://localhost:8501
Grafana: http://localhost:3000
Prometheus: http://localhost:9090

Kubernetes and Helm demo

helm upgrade --install ecommerce-conversion-platform \
  helm/ecommerce-conversion-platform \
  --namespace ecommerce-mlops \
  --create-namespace

The secure wrapper manages the local Kubernetes demo and generates a local Grafana password:

./E-Commerce-Conversion-Intelligence-Platform start
./E-Commerce-Conversion-Intelligence-Platform password
./E-Commerce-Conversion-Intelligence-Platform stop

Quick Start

1. Clone the repository

git clone https://github.com/RuturajM31/E-Commerce-Conversion-Intelligence-Platform.git
cd E-Commerce-Conversion-Intelligence-Platform

2. Create a Python 3.10 environment

python3.10 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip

3. Install the project

python -m pip install -r requirements.txt

4. Launch Streamlit

python -m streamlit run app/Executive_Overview.py

5. Run tests

python -m pytest -q

Optional Isolated MLOps Environments

The main application remains lightweight. MLflow and Evidently are intentionally isolated.

MLflow validation environment

python3.10 -m venv .venv-mlflow
source .venv-mlflow/bin/activate
python -m pip install -r requirements-mlflow.txt
python -m pytest tests/test_mlflow_bridge.py -q

Evidently validation environment

python3.10 -m venv .venv-evidently
source .venv-evidently/bin/activate
python -m pip install -r requirements-evidently.txt
python -m pytest \
  tests/test_evidently_bridge.py \
  tests/test_drift_summary.py \
  -q

Reproducibility and Security

The project uses explicit controls to make results repeatable and repository operations safer.

Dependency strategy

File	Purpose
`requirements.txt`	Full validated project environment
`requirements-app.txt`	Lightweight application runtime
`app/requirements.txt`	Streamlit Community Cloud runtime
`requirements-mlflow.txt`	Isolated MLflow environment
`requirements-evidently.txt`	Isolated Evidently environment
`requirements-ci.txt`	CI-only quality and security tools

Security and reproducibility controls

exact dependency pinning;
pip check compatibility validation;
pip-audit vulnerability scanning;
repository secret scanning;
CI-only disposable fixtures for large local artifacts;
smoke training from sample data;
environment provenance recording;
model serialization and loading tests;
read-only mounts for data, reports, and models in Docker Compose;
generated local Grafana password for the Kubernetes demo;
no committed production credentials.

Repository Structure

E-Commerce-Conversion-Intelligence-Platform/
├── .github/
│   └── workflows/
│       └── ci.yml
├── .streamlit/
│   └── config.toml
├── app/
│   ├── Executive_Overview.py
│   ├── app_utils.py
│   ├── requirements.txt
│   ├── pages/
│   │   ├── 1_Visitor_Intent_Predictor.py
│   │   ├── 2_Batch_Scoring.py
│   │   ├── 3_Model_Benchmark_Selection.py
│   │   ├── 4_Business_KPI_Forecasting.py
│   │   ├── 5_Anomaly_Outlier.py
│   │   ├── 6_Monitoring_Drift_Health.py
│   │   ├── 7_MLOps_Architecture.py
│   │   ├── 8_ML_Validation_Evidence.py
│   │   └── 9_Customer_Segmentation_Journey.py
│   └── ui/
├── src/
│   ├── anomaly/
│   ├── config/
│   ├── data/
│   ├── features/
│   ├── forecasting/
│   ├── models/
│   ├── monitoring/
│   ├── segmentation/
│   └── visualization/
├── data/
│   ├── raw/
│   ├── processed/
│   └── sample/
├── models/
│   ├── trained/
│   └── metadata/
├── reports/
│   ├── figures/
│   ├── final/
│   ├── monitoring/
│   ├── qa/
│   └── tables/
├── monitoring/
│   ├── alertmanager/
│   ├── blackbox/
│   ├── grafana/
│   ├── metrics_cache/
│   ├── prediction_logs/
│   └── prometheus/
├── helm/
│   └── ecommerce-conversion-platform/
├── k8s/
├── scripts/
├── tests/
├── Dockerfile
├── docker-compose.yml
├── Makefile
├── pyproject.toml
└── README.md

Key Outputs

Output	Purpose
`models/trained/final_champion_model.joblib`	Serialized champion model
`models/metadata/final_champion_metadata.json`	Model identity, threshold, metrics, and provenance
`data/processed/final_champion_visitor_scores.csv`	Visitor-level scores and campaign decisions
`reports/tables/final_true_champion_comparison.csv`	Candidate comparison
`reports/tables/final_true_champion_thresholds.csv`	Threshold trade-offs
`reports/tables/final_true_champion_stability.csv`	Stability evidence
`reports/tables/final_true_champion_sensitivity.csv`	Sensitivity evidence
`monitoring/metrics_cache/ecommerce_metrics_snapshot.json`	Lightweight Prometheus source
`monitoring/prediction_logs/`	Prediction ledger and delayed-label evidence
`monitoring/grafana/dashboards/`	Provisioned Grafana dashboards

Large raw files and trained binary artifacts may be excluded from GitHub. Reproducible scripts and sample-data smoke workflows are included so important outputs can be regenerated and validated.

Visual Intelligence Governance

The final Streamlit enhancement program is controlled by a 204-row visual intelligence matrix.

Status	Count
Verified	177
Conditional	25
Excluded	2
Open	0
Total	204

The matrix covers:

executive intelligence;
batch scoring and campaigns;
model selection and explainability;
KNN similarity evidence;
segmentation and journey intelligence;
forecasting and anomaly investigation;
monitoring and drift;
architecture and governance;
shared UI and UX quality;
validation and audit evidence.

This provides explicit evidence that the final application was reviewed as a governed product rather than assembled as an untracked collection of charts.

Portfolio Value

This project demonstrates evidence across the complete analytics lifecycle.

Skill area	Evidence
Business analytics	Conversion framing, KPI interpretation, campaign prioritization
Data engineering	Event-to-visitor transformation and reproducible feature generation
Machine learning	Rare-event modelling, benchmark selection, tuning, threshold economics
Explainable AI	Feature contributions, local similarity, segment and anomaly context
Customer analytics	Segmentation, journey intelligence, campaign capacity
Forecasting	KPI scenarios, uncertainty, residual diagnostics
Monitoring	Evidently drift, prediction ledger, delayed labels, service metrics
MLOps	MLflow, provenance, model metadata, monitoring snapshots
BI and product design	Multi-page Streamlit executive application
DevOps	Docker Compose, Kubernetes, Helm, GitHub Actions
Security and quality	Ruff, pytest, pip-audit, secret scan, smoke training
Technical communication	Architecture diagrams, documented assumptions, decision evidence

The project is directly relevant to roles such as:

Data Analyst;
BI Analyst;
Product Analyst;
Marketing Analyst;
Analytics Engineer;
Data Scientist;
Machine Learning Engineer;
MLOps Engineer.

Honest Limitations

This is a strong portfolio and local production-style platform, but it is not presented as a fully operated enterprise system.

Current boundaries:

model results are based on a public historical dataset;
campaign uplift has not yet been validated through a live randomized experiment;
Streamlit Community Cloud is a portfolio deployment, not a high-availability enterprise serving layer;
Docker and Kubernetes configurations demonstrate deployment readiness but are not continuously operated cloud infrastructure;
monitoring snapshots must be scheduled in a real production environment;
automated retraining and formal model-promotion approval can be expanded;
access control, audit identity, and centralized secret management would be required for enterprise use.

Roadmap

Add campaign A/B testing and uplift measurement
Add scheduled retraining and champion promotion
Add automated monitoring snapshot refresh
Add production feature-store integration
Add event streaming for near-real-time scoring
Add role-based access control
Add managed cloud storage and model registry
Add service-level objectives and incident runbooks
Add formal data-contract enforcement

Dataset

The project uses the public RetailRocket E-Commerce Dataset:

visitor events;
item properties;
category hierarchy.

Dataset source: RetailRocket recommender system dataset on Kaggle

The original event grain is transformed into a visitor-level decision table suitable for purchase-intent modelling and campaign activation.

Author

Ruturaj Mokashi

Data Analyst & BI Specialist focused on transforming complex data into clear, measurable business decisions.

License

This project is licensed under the MIT License.

From raw visitor behaviour to monitored, explainable, campaign-ready intelligence

Launch App · View Repository · View CI/CD

Name		Name	Last commit message	Last commit date
Latest commit History 101 Commits
.github/workflows		.github/workflows
.streamlit		.streamlit
app		app
data		data
docs		docs
helm/ecommerce-conversion-platform		helm/ecommerce-conversion-platform
k8s		k8s
models		models
monitoring		monitoring
reports		reports
scripts		scripts
src		src
tests		tests
.dockerignore		.dockerignore
.env.example		.env.example
.gitignore		.gitignore
Dockerfile		Dockerfile
E-Commerce-Conversion-Intelligence-Platform		E-Commerce-Conversion-Intelligence-Platform
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
docker-compose.yml		docker-compose.yml
pyproject.toml		pyproject.toml
requirements-app.txt		requirements-app.txt
requirements-ci.txt		requirements-ci.txt
requirements-evidently.txt		requirements-evidently.txt
requirements-mlflow.txt		requirements-mlflow.txt
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation

E-Commerce Conversion Intelligence Platform

End-to-end machine learning, business intelligence, and MLOps for purchase-intent prediction

Executive Summary

Project at a Glance

Why This Project Matters

Business Capabilities

End-to-End Platform Architecture

Data Engineering Pipeline

Core visitor features

Machine Learning Strategy

Model lifecycle

Champion decision

Streamlit Business Application

Application pages

Campaign Intelligence

Explainability and Similarity Intelligence

MLflow Experiment Tracking Architecture

MLflow responsibilities

Evidently Drift and Monitoring Architecture

Evidently responsibilities

Delayed Labels and Production Performance

Monitoring and Observability Stack

Monitored areas

Monitoring snapshot pattern

Docker Compose Architecture

Docker services

CI/CD and Quality Gates

CI controls

Deployment Paths

Streamlit Community Cloud

Docker Compose

Kubernetes and Helm demo

Quick Start

1. Clone the repository

2. Create a Python 3.10 environment

3. Install the project

4. Launch Streamlit

5. Run tests

Optional Isolated MLOps Environments

MLflow validation environment

Evidently validation environment

Reproducibility and Security

Dependency strategy

Security and reproducibility controls

Repository Structure

Key Outputs

Visual Intelligence Governance

Portfolio Value

Honest Limitations

Roadmap

Dataset

Author

License

From raw visitor behaviour to monitored, explainable, campaign-ready intelligence

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages