Depth first. Nothing hidden.

The problem

Outdoor heat exposure is not fully represented by air temperature alone. Workers and supervisors need localized Wet Bulb Globe Temperature information, but the underlying forecast data and technical definitions are difficult to access and interpret.

My role

As Junior AI & Data Scientist, I developed and validated Python forecast pipelines using NOAA NBM data, integrated hourly WBGT outputs into the visualization workflow, and helped build the bilingual AI assistant and full-stack product.

System approach

• Download and process structured forecast data with Herbie and xarray.
• Calculate and validate geospatial heat-risk outputs for hourly map updates.
• Render location-aware risk layers in a mobile-focused interface.
• Use a specialized assistant to explain indices and product usage in English and Spanish.

Outcome

The work produced a publicly documented SDSU research product and supported peer-reviewed research on its system architecture, modeling framework and user-centered evaluation.

The problem

NOAA’s real-time mesoscale analysis is scientifically valuable, but its source archive is enormous and its file formats are not designed for casual exploration in a web browser.

My contribution

I contributed across the stack: environmental data processing, API design, geospatial conversion, React and Deck.gl visualization, Docker deployment and the physical Raspberry Pi infrastructure.

System approach

• Acquire hourly RTMA files and process them with xarray, rioxarray and SciPy.
• Serve compact JSON and PNG payloads through FastAPI and ORJSON.
• Render interactive layers with React and Deck.gl.
• Run containerized services from an external SSD with Cloudflare Tunnel access.
• Connect a LLaMA 3.2 Vision assistant to screenshots and map context.

What it proves

The project demonstrates end-to-end ownership under real hardware constraints. The source archive exceeds 91 TB; instead of pretending one device can hold it all, the system keeps a focused rolling cache and serves the slice a user needs.

The challenge

Delays and cancellations are imbalanced outcomes affected by time, route, carrier and weather. A useful comparison required careful merging, feature engineering and metrics that did not hide minority-class performance.

My contribution

I worked within the project team on the machine-learning workflow, analysis and communication, then presented the research publicly as a WUSS student scholar.

Modeling approach

• Merge aviation and weather sources and reduce the analysis set to 1.25M rows.
• Compare logistic regression, trees, random forests, boosting, KNN and neural networks.
• Evaluate accuracy alongside macro and weighted F1 to make imbalance visible.
• Package the analysis in an interactive Streamlit application.

Result, stated precisely

XGBoost delivered approximately 0.68 accuracy, 0.70 weighted F1 and 0.62 macro F1. The gap between weighted and macro performance is part of the finding—not a number to hide.

The question

How do traffic patterns vary alongside population growth, geographic location and economic conditions across a state as large and diverse as California?

My contribution

I cleaned and analyzed population data, contributed to the combined dataset and visualizations, and helped design and write the public-facing project website.

Analytical approach

• Normalize and combine Caltrans traffic records with Census population and income.
• Explore county, coordinate and time patterns in Tableau, Python and R.
• Compare OLS and Random Forest models under different feature selections.
• Document multicollinearity and the limits of interpreting observational relationships.

What I would improve now

Several high-scoring model variants relied on closely related traffic measures or geographic coordinates. Today I would add spatial cross-validation and stricter leakage controls before treating predictive performance as generalizable.