Projects

Selected work in data analysis, statistical modeling, machine learning, experimentation, and applied AI.

Marketplace Operations Analytics

Operational analytics on a multi-table e-commerce marketplace, focused on building trustworthy metrics and diagnosing fulfillment, seller, category, and geographic performance.

  • Integrated relational marketplace data using Python, SQL, and PostgreSQL.
  • Validated referential integrity, join coverage, fan-out, missingness, and metric consistency before analysis.
  • Investigated fulfillment performance and operational differences across sellers, categories, and geographies.
  • Built reusable KPI datasets for operational analysis and Tableau reporting.
  • Python
  • SQL
  • PostgreSQL
  • Tableau

Retail Media Measurement & Decision Platform

Experimentation and measurement on synthetic retail-media data, built to separate attributed performance from causal incremental impact.

  • Analyzed randomized treatment-control experiments to estimate incremental conversion, orders, and revenue.
  • Quantified uncertainty using confidence intervals and stratified bootstrap estimates.
  • Built campaign- and segment-level measurement datasets using consistent experiment-window definitions.
  • Developed reproducible Python and SQL workflows for campaign performance and incrementality analysis.
  • Python
  • SQL
  • PostgreSQL
  • Experimentation

Vector RAG Backend

Retrieval-augmented generation backend designed to retrieve relevant source material and produce grounded answers with traceable citations.

  • Built document ingestion, chunking, embedding, and vector-retrieval workflows.
  • Implemented PostgreSQL and pgvector retrieval for semantically relevant source selection.
  • Exposed retrieval and generation workflows through a FastAPI service.
  • Evaluated retrieval behavior and designed responses around source grounding and citation traceability.
  • Python
  • FastAPI
  • PostgreSQL
  • pgvector
  • RAG

Search & Ranking Platform

Product-search relevance project comparing lexical retrieval, learning-to-rank, and semantic relevance approaches on query-level candidate sets.

  • Established TF-IDF and BM25 baselines and evaluated rankings with NDCG@10, Recall@10, and MRR.
  • Built query-product relevance features covering lexical similarity, token coverage, brand matching, and length signals.
  • Trained an XGBoost learning-to-rank model with query-level grouping and strict query-based train/validation separation.
  • Extended the ranking pipeline with embedding-based semantic relevance features.
  • Python
  • XGBoost
  • BM25
  • TF-IDF
  • Embeddings