Projects & Skills

A governed analytical question-answering system over structured data. The model never executes SQL and never directly calculates a published value; it compiles questions into typed analytical contracts, executes them locally over DuckDB, and verifies every result before publishing, with full provenance attached.

A LangGraph research pipeline that addresses hallucination and citation fabrication in single-model workflows. A generative model writes claims and quotes; a separate NLI entailment classifier and publication gate decide what gets published. Everything else is withheld. Blinded self-review returned precision 1.00 with zero unsupported claims published.

A budget-constrained alert evaluation framework for anti-money-laundering transaction monitoring, built on IBM's AMLworld benchmark. Reports precision@k under a fixed daily alert budget with a random-ranker null, an attainable ceiling, and content-addressed provenance, so every published number traces to a committed artifact.

Event-driven AWS ELT pipeline with Snowflake, dbt Medallion architecture, SCD Type 2 snapshots, and a 7-task fail-fast Airflow DAG with 64 automated data quality tests.

End-to-end computational biology pipeline comparing a CNN against three RNA foundation models (RNA-FM, RNABERT, SpliceBERT) across 16 RNA-binding proteins — SpliceBERT achieved mean AUROC 0.893 and transferred to ClinVar variant-effect analysis with AUROC 0.837 on noncoding variants, without any disease-label training.

Distributed PySpark pipeline on Databricks ingesting 150M+ records across 5 sources, training 4 MLlib regression models, and benchmarking parallelism scaling from single-node to multi-node Spark – achieving 2.73x speedup.

Near real-time wildfire monitoring pipeline using AWS Lambda, S3, and Glue, feeding live geospatial Tableau dashboards that auto-refresh every 3 hours – reducing manual tracking by 85%.

ADF-orchestrated Databricks pipeline with parameterized incremental loads, star schema in Delta Lake with SCD Type 1, Unity Catalog governance, and high-water mark ingestion.

Hybrid MySQL + MongoDB system managing patient data, treatments, clinical trials, researchers, and outcomes – with Python analytics covering demographics, adverse events, and drug utilization.

High-performance PostgreSQL analytics across 1M+ Apple retail transactions – solving 23 business problems with query optimizations delivering up to 95% faster execution and 54% performance gains.

Python-based EDA and ARIMA time-series forecasting on LA crime data from 2020–2024, identifying vehicle theft as the top crime category and projecting a 14.89% crime reduction through late 2025.