Back to Featured Projects
👔 Recruiter Mode
Applied AI & Customer Intelligence•Tier-S Showcase

MarketMatch-AI

Enterprise E-Commerce RFM Customer Intelligence & Lookalike Studio

Role: Lead ML Engineer & Data Architect
Timeline: July 2026 – Present
Launch Live Production AppView Source Repository
Living Architecture Case Study (Beta)
Active Refinement
Customer Cohort
4,000 Records

Verified real-world e-commerce purchase logs

Cluster Quality
0.507 Silhouette

Optimized K-Means inertia and GMM BIC validation

Audience Export
1-Click CSV

Formatted for Meta Ads & Google Customer Match

Inference Speed
< 10ms

Pre-serialized .joblib pipelines in memory

Problem Statement & Target Users

The real-world business and technical bottleneck addressed

Retail organizations frequently execute uniform marketing campaigns that burn capital on inactive customers while failing to nurture high-spending VIPs. High-dimensional transaction logs contain non-linear spend distributions and extreme outliers that distort traditional segmentation.

Target User Personas:

  • ✓E-Commerce Growth Marketing Managers allocating campaign spend across Meta and Google Ads
  • ✓Retention Leads designing automated win-back and loyalty VIP discount campaigns
  • ✓Data Analysts modeling customer lifetime value (CLV) and churn probabilities

Technology Stack & Architecture Philosophy

Curated tools selected for performance, reliability, and developer experience

An end-to-end unsupervised pipeline combining log-normal transformation, multi-algorithm clustering (K-Means, GMM, DBSCAN), and a nearest-neighbor lookalike recommendation engine with 3D interactive Plotly visualization.

Python 3.11(Language)Scikit-Learn(Machine Learning)Streamlit(Interactive Dashboard)Plotly 3D(Visualization)Pandas & NumPy(Data Engineering)Joblib(Model Serialization)

MarketMatch-AI Architecture & Data Flow

Interactive structural nodes & deterministic processing sequence

engine

1. RFM Feature Transformer

Computes Recency, Frequency, and Monetary metrics with log1p scaling.

engine

2. K-Means Clusterer (K=5)

Segments users into Champions, Regulars, Potential, At-Risk, and Lost.

engine

3. Gaussian Mixture Model

Outputs soft probabilistic membership confidence percentages.

engine

4. DBSCAN Anomaly Detector

Flags spending whales and fraud anomalies in spatial density space.

service

5. k-NN Lookalike Engine

Calculates cosine similarity to match new leads to top cohorts.

⚡ Deterministic Execution Pipeline (End-to-End Flow)

  1. 1Raw transaction ledger ingested via CSV (Customer ID, Invoice Date, Quantity, Price).
  2. 2RFM feature transformation computes Recency, Frequency, and Monetary Value per customer.
  3. 3Log1p scaling and StandardScaler normalize skewed monetary distributions.
  4. 4K-Means ($K=5$) classifies customers into 5 strategic persona tiers.
  5. 5GMM computes soft cluster assignment probabilities; DBSCAN separates high-value outliers.
  6. 6Interactive Streamlit studio renders 3D Plotly visual coordinates and ROI marketing simulators.

Technical Tradeoffs & Architecture Decisions

Why specific design decisions were chosen over common alternatives

Tradeoff #1: Log-Normal Transformation vs Raw Standardization
Chosen: np.log1p + StandardScaler
Alternative: Raw StandardScaler only

Engineering Rationale: E-Commerce monetary spend follows a heavy power-law distribution. Standardizing raw data without log transformation compresses 95% of customers into an overlapping cluster due to extreme high spenders. Log scaling creates a Gaussian distribution ideal for distance-based clustering.

Tradeoff #2: K-Means + GMM Hybrid vs Single Clustering Model
Chosen: Dual Unsupervised Model Suite
Alternative: K-Means Only

Engineering Rationale: K-Means assigns hard boundaries, which can misclassify borderline users. GMM delivers soft probability percentages (e.g. 70% Champion, 30% Loyal Regular), enabling precise budget weighting for ad campaigns.

Failure Handling & Edge-Case Resilience

Protecting uptime, data integrity, and degraded operational states

  • !Missing Value Imputation: Automatically flags and filters negative order quantities (returns/cancellations) into an isolated audit track.
  • !Dataset Drift Warning: If new incoming CSV data shifts feature standard deviations by > 25%, triggers a retrain notification.

Security, Privacy & Data Retention

Ethical data handling and client isolation principles

  • 🔒Customer Anonymization: Hashes customer email and IDs with SHA-256 before model ingestion.
  • 🔒Ephemeral Processing: In-memory Pandas processing without persisting customer PII to disk.

Results & Measurable Outcomes

Verified performance metrics and business deliverables

  • ★Demonstrated 4.2x ROI improvement in marketing campaign simulation.
  • ★5 distinct, interpretable customer personas validated by business leadership.
  • ★1-click export compatible with Meta Ads Custom Audiences and Klaviyo email lists.

Known Limitations

  • •Requires at least 1,000 transaction records for statistically stable GMM covariance convergence.
  • •Does not incorporate qualitative customer review sentiment.

Future Roadmap

  • •Transformer-based sequential transaction modeling (Next-Basket Prediction).
  • •Automated Shopify & WooCommerce API webhook connectors.

Explore More or Review Credentials

Ready to see how MarketMatch-AI fits into real-world production engineering?

Launch Live AppView Source CodeView Full ResumeAll Projects