Data analytics projects

2.8M+ traffic records turned into Tableau & Streamlit KPIs for non-technical stakeholders

Analyzed 2.8M+ public traffic enforcement records to surface trends, patterns, and KPIs for non-technical stakeholders. Built Python ETL, Tableau, and Streamlit dashboards for safety and enforcement insights.

Serverless ETL (S3 → Glue → Athena → dashboards) at ~$1–2/month with Parquet cost savings

End-to-end serverless ETL for e-commerce sales data—from ingest to SQL analytics and dashboards—at ~$1–2/month. S3 + Lambda + Glue into Parquet, Athena SQL, QuickSight/interactive dashboards, with data-quality checks.

4M-review NLP pipeline with locked holdout — F1 ~0.906, ROC-AUC ~0.966

Group capstone on binary sentiment over a 4M Amazon Reviews corpus (3.6M train / 400k locked test). Compared linear models, Random Forest, MLP, and DistilBERT under strict leakage controls—TF–IDF + linear baseline reached F1 ~0.906 / ROC-AUC ~0.966.

30-day readmission models across 130 U.S. hospitals + interactive Shiny dashboard

Interpretable ML to predict 30-day hospital readmissions across 130 U.S. hospitals—logistic regression, CART, and Random Forest—with an R Shiny dashboard for clinical/analytics stakeholders.

Isolation Forest anomaly detection with a live Streamlit monitoring dashboard

Unsupervised anomaly detection for mobile-money-style transactions using Isolation Forest, with behavioral/geographic/temporal features and a live Streamlit monitoring dashboard.

Neo4j + vector GraphRAG research assistant with grounded citations and streaming UI

Built a graph- and vector-grounded research assistant over a scientific PDF corpus with route-aware retrieval, streamed citations, and a production chat UI—combining Neo4j knowledge graphs, vector search, and grounded LLM answers.

87k+ BOM relationships; PageRank and Leiden for supply-chain risk triage

Group capstone modeling a multi-tier automotive supply network in Neo4j (87k+ BOM links, facilities, products), then using graph algorithms—PageRank for critical chokepoints and Leiden community detection for modular risk views—to prioritize disruption triage over expensive recursive SQL patterns.

Serverless sentiment pipeline on AWS Free Tier with 75–80% accuracy

Built a production-ready sentiment analysis system for customer feedback analysis using AWS Free Tier services. Implemented serverless architecture with Lambda, DynamoDB, S3, and API Gateway, creating a cost-effective solution that processes thousands of feedback items with 75-80% accuracy.

Relational-to-graph Chinook model with constraint-first CSV ingestion

Modeled a Chinook-style relational music dataset as a Neo4j property graph—designing nodes and relationships, loading CSVs in dependency order with constraints and MERGE, and validating the schema with Cypher queries.

Interactive Streamlit fitness analytics and progress visualization

Developed an interactive fitness tracking application using Streamlit and Python. Features include workout analytics, progress visualization, and personalized recommendations based on user data patterns.

FAERS-style graph analytics from Cypher EDA through purposeful GDS workflows

Explored a FAERS-style healthcare knowledge graph with Cypher EDA, analytical aggregations, and Graph Data Science projections—using Python + Neo4j driver + pandas to turn structural and similarity results into practical analytical insights.

Hybrid graph recommender combining collaborative and content signals with GDS embeddings

Built a graph-based movie recommendation system combining collaborative rating edges, content relationships (genres, directors), and GDS algorithms—FastRP embeddings, kNN similarity, and Louvain communities—to handle sparse user-item data with explainable recommendations.