Portfolio

Sample Work Projects

Texas Afghan Support Program — Geospatial Analytics

Organization: Cherokee Federal/ Texas Office for Refugees (TXOR)

Overview: Designed and developed an end-to-end Python geospatial analytics application to examine program participation, rental assistance distribution, and Financial Wellness Course engagement across major Texas metropolitan areas, including Austin, Dallas, Fort Worth, Houston, and San Antonio.

All program values are synthetic and intentionally perturbed for portfolio use. No client-level or personally identifiable information is included.

Key Technical & Operational Features:

  • Geospatial & Quantile Analysis: Built interactive choropleth layers using Folium, Leaflet, and GeoJSON to visualize household enrollment, participant distribution, rental assistance, and service engagement at the ZIP-code level.
  • Exploratory Data Analysis: Developed a Jupyter-based analytical workflow examining metro-area differences, geographic concentration, monthly program trends, normalized measures, correlations, and month-over-month changes.
  • Interactive Data Exploration: Implemented detailed ZIP-level tooltips and click/tap detail cards, ZIP-code search, metro-area quick navigation, dynamic legends, optional heat-map and Top-10 overlays, and an integrated Map Guide.
  • Data Validation & Reproducibility: Created automated quality checks covering schema integrity, cumulative enrollment logic, rental-assistance reconciliation, derived measures, and geographic rankings. A reproducible Python pipeline regenerates analytical outputs and the interactive map from the source data.
  • Continuous Integration & Deployment: Configured pytest and GitHub Actions to automatically validate the dataset, rebuild the analytical application, and deploy the live map through GitHub Pages.

Skills & Tools Used: Python â€¢ pandas â€¢ NumPy â€¢ Folium â€¢ Leaflet â€¢ GeoJSON â€¢ Jupyter â€¢ Matplotlib â€¢ pytest â€¢ GitHub Actions â€¢ Geospatial Analysis â€¢ Exploratory Data Analysis â€¢ Data Visualization

Public Program Compliance Monitoring & Risk Analytics

Regulatory Compliance Analytics & Predictive Risk Modeling

Overview: Developed an end-to-end compliance analytics system for a synthetic state-administered provider program. The project translates regulatory monitoring requirements into reproducible SQL controls, KPI reporting, exception detection, remediation tracking, and predictive risk prioritization.

Key Features & Technical Execution

  • SQL Compliance Monitoring: Built auditable rules for reporting timeliness, documentation completeness, complaint activity, unresolved corrective actions, and repeat violations.
  • Provider Scorecards: Aggregated provider-level compliance history into review-ready scorecards with exception counts, material findings, complaint rates, and remediation performance.
  • Corrective-Action Aging: Created a normalized remediation register with opened, due, and closure dates to identify overdue, late-closed, and high-severity unresolved actions.
  • Threshold Sensitivity: Evaluated how alternative compliance thresholds affect alert volume and analyst workload.
  • Severity Tiers: Developed transparent Routine, Watch, High, and Critical review tiers using deterministic controls and explicit reason codes.
  • Temporal Machine Learning: Built models to predict material compliance findings within the next 90 days using expanding time-based validation with a three-month embargo.
  • Model Comparison: Compared deterministic rules, logistic regression, and gradient boosting using ROC-AUC, average precision, Brier score, and log loss.
  • Feature Ablation: Reduced the final predictive specification to a 10-feature compliance-only model after showing that provider type, region, and lower-value operational features reduced out-of-time performance.
  • Responsible Model Design: Excluded provider type and region from predictive scoring while retaining them for segment-level performance auditing.
  • Decision Support: Evaluated calibration, lift, risk concentration, model-versus-rule overlap, coefficient interpretation, and review-capacity tradeoffs.
  • Reproducibility: Added automated testing, GitHub Actions workflows, analytical metadata, model contracts, and repository-integrity validation.

Selected Results

  • 250 synthetic regulated providers
  • 6,000 provider-month observations
  • 1,046 corrective actions
  • 59 overdue open corrective actions
  • Preferred model ROC-AUC: 0.728
  • Average precision: 0.384
  • Highest-risk 10% captured 30.3% of subsequent material findings
  • Top-decile lift: 3.03×

Outcome: Produced a portfolio-ready compliance analytics workflow that demonstrates how regulatory requirements can be translated into SQL monitoring, operational reporting, remediation analysis, and explainable predictive risk prioritization. All data are synthetic, and model outputs are designed to support human review rather than automate compliance determinations.

SIPP Economic Stability & Poverty Transitions

Longitudinal Poverty Transition Analysis

Overview: Developed a longitudinal public-policy analysis using the U.S. Census Bureau Survey of Income and Program Participation (SIPP) to measure annual poverty transitions and build prospective models of poverty entry and exit. The project links respondents across the 2024 and 2025 SIPP releases, preserving temporal ordering between baseline economic conditions and following-year outcomes.

Key Features & Technical Execution

  • Longitudinal Cohort Construction: Linked respondents across consecutive SIPP releases using SSUID + PNUM, producing a validated analytical cohort of 15,034 adults.
  • Complex-Survey Inference: Used Census two-year longitudinal person weights and all 240 replicate weightswith Fay’s BRR variance estimation.
  • Poverty Transition Analysis: Measured movement between stable-above-poverty, entered-poverty, exited-poverty, and persistent-poverty states.
  • Prospective Machine Learning: Built separate poverty-entry and poverty-exit models using baseline characteristics only, preventing future information from leaking into model inputs.
  • Household-Grouped Validation: Kept members of the same household on the same side of each train/test split.
  • Policy-Facing Model: Developed a validated 10-feature weighted logistic regression using education, employment, income, housing, and SNAP characteristics.
  • Model Comparison: Evaluated logistic regression and gradient boosting using ROC-AUC, average precision, Brier score, log loss, and calibration.
  • Feature Ablation: Tested the predictive value of poverty ratio, income, assets, safety-net variables, housing characteristics, and demographics.
  • Responsible AI: Found that removing demographic predictors slightly improved overall performance. Age, sex, race, Hispanic origin, nativity, and citizenship were therefore excluded from prediction and retained only for subgroup auditing.
  • Risk Diagnostics: Evaluated risk concentration and subgroup performance to assess how the model behaves beyond aggregate accuracy.
  • Reproducible Engineering: Built source verification, schema validation, streaming data processing, automated testing, integrity checks, and GitHub Actions workflows.

Selected Results

  • Longitudinal cohort: 15,034 adults
  • Poverty-entry rate among baseline nonpoor adults: 4.05%
  • Poverty-exit rate among baseline poor adults: 48.43%
  • Preferred model ROC-AUC:
    • Poverty entry: 0.720
    • Poverty exit: 0.712
  • Highest-risk 10% of the primary entry holdout contained 32.1% of observed poverty entries

Outcome: Produced a reproducible end-to-end policy analytics workflow combining longitudinal survey methodology, complex-survey inference, prospective machine learning, model governance, and responsible feature selection.

2025 NHIS Healthcare Access Analysis

Healthcare Access Analysis

Overview: Developed a reproducible Python analysis examining cost-related delayed medical care among U.S. adults using the 2025 National Health Interview Survey. The project combines complex-survey inference, multiple imputation, subgroup analysis, and predictive modeling to evaluate disparities in healthcare access.

Key Features & Technical Execution

  • Complex-Survey Analysis: Estimated population prevalence and subgroup differences using NHIS survey weights, strata, and primary sampling units.
  • Variance Estimation: Used Taylor-linearized standard errors and confidence intervals to preserve the NHIS survey design.
  • Multiple Imputation: Retained and analyzed all 10 official NHIS income imputations, pooling poverty-related estimates using Rubin’s rules rather than relying on a single imputed dataset.
  • Healthcare Access Outcome: Examined whether adults reported delaying medical care because of cost.
  • Subgroup Analysis: Evaluated access disparities across demographic, socioeconomic, insurance, and health-related characteristics.
  • Predictive Modeling: Built models using demographic, socioeconomic, healthcare-access, and health-status characteristics to distinguish adults reporting cost-related delayed care.
  • Model Evaluation: Compared model performance using discrimination and probability-based metrics rather than relying on classification accuracy alone.
  • Survey-Aware Interpretation: Kept descriptive population inference separate from predictive modeling so national estimates were not confused with model outputs.
  • Reproducibility: Built a documented analysis pipeline with structured notebooks, validated inputs, reusable code, and GitHub-ready documentation.

Outcome: Produced a rigorous healthcare-access analysis demonstrating complex-survey methods, multiple imputation, subgroup analysis, and predictive modeling using a nationally representative federal health dataset.

Texas Refugee Medical Assistance (RMA)

Longitudinal Geospatial Analysis (FY2019–FY2024)

Organization: U.S. Committee for Refugees and Immigrants (USCRI) 

Overview: Developed a Python-based longitudinal and geospatial analytics project examining how synthetic Refugee Medical Assistance enrollment patterns change across Texas from FY2019 through FY2024. The project combines a complete six-year ZIP-level panel, demographic and concentration analysis, data validation, and an interactive time-driven mapping application.

All program values are synthetic and intentionally perturbed for portfolio use. No client-level or personally identifiable information is included.

Key Features & Technical Execution

  • Longitudinal & Geographic Analysis: Analyzed statewide growth, geographic expansion, city trajectories, ZIP persistence and emergence, and country-of-origin composition across six fiscal years.
  • Interactive Geospatial Application: Built a time-driven map with fiscal-year selection, enrollment-intensity and dominant-country modes, ZIP search, statewide summaries, dynamic rankings, and six-year ZIP trend visualizations.
  • Demographic & Concentration Analysis: Evaluated changes in country-of-origin composition using enrollment shares and the Herfindahl-Hirschman Index (HHI), providing context beyond raw enrollment counts.
  • Reproducible Data Pipeline & Validation: Created an automated Python workflow for data processing, longitudinal analysis, application generation, and data-quality testing, with GitHub Actions used for continuous integration and deployment.
  • Technologies: Python, pandas, NumPy, Leaflet, GeoJSON, Matplotlib, Jupyter, pytest, GitHub Actions

Refugee Resettlement Capacity & Resource Planning

Overview: A reproducible decision-support platform for forecasting refugee-arrival demand, testing service-capacity scenarios, and allocating hypothetical service resources under transparent constraints.

What the project does

  • Ingests and validates U.S. Refugee Processing Center arrival reports across changing XLS, XLSX, and PDF formats.
  • Maintains a validated national monthly arrival series for FY2021–FY2025 and a reconciled FY2024 state-month panel.
  • Benchmarks naive, seasonal-naive, exponential-smoothing, SARIMA, and gradient-boosting forecasts with leakage-safe rolling-origin validation.
  • Converts a reference forecast into explicit lower, baseline, and higher demand scenarios.
  • Evaluates staffing capacity, service coverage, surplus capacity, and unmet demand.
  • Allocates a limited staffing pool across state planning units with max-min fair linear programming.
  • Adds budget, FTE, state-ceiling, and minimum-coverage constraints.
  • Exposes the pipeline through FastAPI and an interactive scenario-planning interface.
  • Packages the application with Docker and validates the production image in GitHub Actions.

Texas Refugee Service Access Atlas

Overview: Geospatial decision-support platform for analyzing scheduled public-transit access to refugee-relevant services in Dallas and Tarrant Counties, Texas.

The project combines Census neighborhood context, verified service-provider locations, DART and Trinity Metro GTFS schedules, multimodal routing, service-specific accessibility, transparent gap classification, and facility-location optimization.

Core question: Which Dallas–Fort Worth neighborhoods face the greatest barriers to reaching refugee-relevant services by public transit?

What the project does

The atlas allows users to:

  • shade census tracts by modeled nearest-provider travel time
  • compare foreign-born population count and share, no-vehicle households, and poverty context
  • toggle 30-, 45-, and 60-minute accessibility thresholds
  • compare frequency-agnostic and schedule-aware transit scenarios
  • filter access by service category
  • inspect service-gap classifications
  • view verified provider locations
  • compare facility-location planning scenarios
  • read a plain-language summary of the selected service/time scenario
  • compare frequency-agnostic network potential with scheduled 45-minute access
  • compare current coverage with modeled one-site intervention coverage
  • export current tract results, gap priorities, provider inventory, and facility scenarios as CSV

The application is deployed as a Dockerized FastAPI service on Render and reads directly from the validated processed outputs in this repository. GitHub Actions now validates the production Docker image and checks that the live Render service is running the validated commit or a newer descendant.

U.S. Immigration Policy Intelligence

Federal Statute & Semantic Search Platform

Overview: Developed a full-stack policy intelligence platform for analyzing U.S. federal immigration activity published in the Federal Register from 2017–2026. The project combines API-based data ingestion, PostgreSQL, SQL analytics, natural language processing, semantic retrieval, and a deployed FastAPI web application to examine rule-making patterns and make federal immigration-policy documents easier to search and explore.

Key Features & Technical Execution:

  • API-Driven Data Engineering: Built a reproducible Federal Register ingestion pipeline covering USCIS, CBP, ICE, and EOIR, with auditable inclusion rules, raw-data preservation, validation, deduplication, and support for documents associated with multiple agency scopes.
  • PostgreSQL & SQL Analytics: Designed a normalized relational database preserving document, agency, CFR, Regulation Identifier Number (RIN), and corpus-evidence relationships. Developed reusable SQL analyses examining activity over time, agency patterns, frequently referenced CFR provisions, RIN reuse, and proposed-rule-to-final-rule timelines.
  • Natural Language Processing: Applied TF-IDF, cosine similarity, K-means clustering, model-selection diagnostics, and representative-document analysis to a curated 277-document substantive-policy corpus, identifying recurring themes such as asylum and removal policy, Temporary Protected Status, immigration-court procedures, employment authorization, and regulatory fees.
  • Semantic Retrieval & Evaluation: Implemented MiniLM semantic search using FastEmbed and ONNX Runtime and compared it with a TF-IDF baseline. Semantic retrieval achieved 0.844 mean precision@5 compared with 0.733 for TF-IDF, while both approaches achieved a 100% hit rate@5 in the exploratory benchmark.
  • Interactive Policy Research Application: Developed and deployed a FastAPI web application supporting natural-language search with agency, year, document-type, and policy-topic filters. Search results include abstract snippets and similarity scores, while document-detail pages provide Federal Register citations, CFR references, RINs, effective dates, comment deadlines, official source links, and related documents.
  • Reproducibility & Deployment: Created automated database, SQL, NLP, retrieval, API, Docker, and public-deployment validation workflows using pytest and GitHub Actions. Containerized the application with Docker, pinned and locked dependencies, embedded the ONNX model in the deployment image, and deployed the application to Render with automated live smoke testing.
  • Technologies: Python, PostgreSQL, SQL, pandas, NumPy, scikit-learn, FastEmbed, ONNX Runtime, FastAPI, HTML/CSS/JavaScript, Docker, pytest, GitHub Actions, Render
A laptop sits on a desk, displaying a software interface with a blue and purple data network visualization on the screen.

Texas Statewide RMA Enrollment Analytics & Interactive Heat Mapping Framework

Organization: U.S. Committee for Refugees and Immigrants (USCRI) 

Overview: Initiated and led a data analytics initiative to monitor statewide Refugee Medical Assistance (RMA) enrollment across Texas. After designing a foundational Power BI enrollment dashboard, expanded the effort into a quarterly reporting and geographic analytics framework to deliver actionable insights to leadership, resettlement agencies, and state partners.

All values are synthetic and intentionally perturbed for portfolio use. No client-level or personally identifiable information is included.

Key Deliverables & Actions:

  • Statewide Analytics & Reporting: Authored the FY25 Q1 Statewide RMA Enrollment Report, aggregating data across 20+ resettlement partner locations to evaluate 7,024 client enrollments.
  • Interactive Heat Maps: Built dynamic, HTML/browser-accessible zip-code heat maps to visualize enrollment concentrations across major metropolitan areas (Houston, Dallas-Fort Worth, Austin, San Antonio) and regional hubs.
  • Demographic & Policy Analysis: Analyzed key enrollment trends, tracking status distributions (e.g., Cuban/Haitian Entrants representing 83.4% of enrollments), language preferences, age/gender breakdowns, and conditional enrollment factors (e.g., 65+ age group).
  • Stakeholder Alignment: Translated raw data into strategic visualizations, enabling executive leadership and partner agencies to optimize resource allocation, address compliance bottlenecks, and plan targeted outreach.

Impact: Transformed isolated case data into an interactive intelligence asset. Provided the organization with its first automated, multi-agency tracking framework to guide capacity planning and state-level policy discussions

eRED RMA Application Portal Redesign & User Experience Modernization

Organization: U.S. Committee for Refugees and Immigrants (USCRI) 

Overview: Refined and optimized the Electronic Refugee Initial Determination (eRED) web portal to streamline case processing for partner agency case managers. Focused on restructuring application elements to align with actual operational workflows and embedding policy guidance directly into the interface.

Key Deliverables & Actions:

  • Intuitive Workflow Restructuring: Reordered application fields within the existing three-step structure to create a natural, efficient logical flow tailored to case managers’ daily entry habits.
  • Interactive Policy Guidance: Integrated field-level tooltips throughout the form, providing instant, on-demand policy clarifications and guidance to reduce submission errors and minimize external policy lookups.
  • Interactive Prototyping: Built a functional HTML/JS prototype incorporating the new field order and tooltip mechanics to demonstrate the user-centric workflow improvements to leadership.

Impact: Created a user-friendly, self-guiding application interface concept that reduced administrative confusion, improved compliance accuracy, and provided a blueprint for future portal iterations.

Statewide Public Assistance Training & Compliance Framework (SNAP/TANF/Medicaid)

Inter-Agency Instructional Deck & Regulatory Compliance Guide for Refugee Service Providers

Organization: Texas Office for Refugees (TXOR)

Overview: Developed and facilitated comprehensive training modules for partner agencies across Texas to standardize benefit application processing, reduce eligibility discrepancies, and navigate complex state regulatory guidelines (HHSC).

Key Portfolio Highlights:

  • Complex Policy Translation: Visualized multi-program eligibility criteria, income threshold calculations, and immigration documentation verification workflows into accessible staff training modules.
  • Operational Accuracy: Built technical guidance covering self-employment declarations (Form H1049), earned income projection formulas, and domicile verification standards.
  • Appeals & Escalation Framework: Designed step-by-step checklists and template language for case managers to submit policy-backed benefit appeals under Texas Works Handbook guidelines.

Deliverables Included: Instructional Deck • Process Flowcharts • Verification Checklists • Sample Appeal Documentation

Skills & Tools: Policy Analysis â€¢ Instructional Design â€¢ Stakeholder Facilitation â€¢ Regulatory Compliance â€¢ SOP & Technical Writing