50+ executives at a 1M+ member health insurer wasted 40h/week manually assembling Excel reports from 4 incompatible regulated sources (RIPS, BDUA, tutelas, clinical records).

Designed a cloud data warehouse integrating all 4 sources with automated ETL pipelines in Python/Airflow. Built geospatial predictive models for demand forecasting. Delivered Power BI dashboards with custom HTML/CSS visuals and live refresh — zero manual intervention.

Eliminated 100% of manual Excel reports. Critical reporting cycle dropped from 3 days to 4 hours. Geospatial models hit 85% accuracy, reshaping resource allocation for high-risk zones.

1,841 temporary disability complaints (Jan–Jun 2025) from a Colombian ARL insurer sat unanalyzed. Manual review was slow, inconsistent, and couldn't surface systemic patterns across claim types.

Built an end-to-end pipeline: NLP text classification of complaint narratives, interpretable ML models (SHAP explainability), an interactive Streamlit dashboard for claims analysts, and a full automation pipeline for future ingestion.

Surfaced hidden complaint clusters and risk patterns invisible to manual review. Delivered an analyst-facing dashboard and reproducible ML pipeline ready for production deployment.

A rural hospital's ER faced severe overcrowding. No data existed on demand patterns, staff allocation was reactive, and patients waited 2–3x longer than benchmarks. Medical liability cases piled up.

Built time-series demand models in Python/SQL to predict ER load by hour and day. Deployed a geospatial telemedicine routing system to redirect low-acuity patients. Set up 12 Power BI dashboards tracking 15 KPIs in real time. Ran forensic data analysis for 90+ medical liability cases.

Emergency wait times cut by 35%. Telemedicine covered 2,000+ rural users, reducing in-person costs by 40%. 90% favorable resolution rate on medical liability cases through data-driven forensic analysis.

Standard bird recognition systems fail to capture how birds actually perceive color — they see in 4 channels (UV included), not 3. Existing datasets and models ignored this entirely.

Trained a multi-class CNN + Vision Transformer (BEiT) pipeline on GCP (PyTorch + OpenCV) that classifies species and simulates tetrachromatic UV vision. HDBSCAN clustering with UMAP projections for unsupervised species grouping. Deployed as a Streamlit app with interactive spectral heatmaps.

First open-source model combining bird classification with tetrachromatic vision simulation. Validated improvements in clustering metrics (Silhouette Score, Calinski-Harabasz). Published as a reproducible ML pipeline.

Demonstrated end-to-end data engineering competency across 5 real-world domains: unified data lakes, YouTube trends, social sentiment, security SIEM logs, and open traffic data.

Built each pipeline with production-grade tooling: Airflow orchestration, Spark for distributed processing, Kafka for real-time streaming, dbt for transformations, Delta Lake for storage, and BERT/Isolation Forest for ML layers. All containerized with Docker Compose.

5 end-to-end pipelines covering batch + streaming + ML inference, deployable via Docker. Covers the full modern data stack from ingestion to Grafana/Kibana dashboards.

Querying NCBI BLAST for genomic sequence analysis required deep bioinformatics knowledge and repetitive manual queries — a barrier for researchers without programming backgrounds.

Built an autonomous AI agent with LangChain + OpenAI API that translates natural language into BLAST queries, parses results, and returns structured insights. Deployed on Streamlit for zero-install access.

Researchers query genomic databases in plain English. Live demo running on Streamlit — no CLI, no code, no bioinformatics background required.