cv

Freelance AI Engineer / Machine Learning Engineer / Data Scientist

Freelance (Remote)

Design, build, and deploy production software, machine learning, and agent-based AI solutions spanning LLM evaluation, autonomous agents, analytics, data engineering, and intelligent automation using Python, SQL, and modern cloud technologies.

• Built and published the AI Fluency Suite (AgentFluent + CodeFluent), open-source tools for evaluating AI agent quality and human-AI collaboration effectiveness
• Shipped production ML services on Google Cloud Run with CI/CD and full test coverage: a sportswear ESG news classifier with a live sustainability scorecard, an e-commerce fraud detection API, and a public U.S. presidential voting data API
• Earlier client work: legal tech analytics, political data analysis, startup analytics and reporting

Strategic Data Analyst

Portland State University Foundation

Completed end-to-end training on Blackbaud CRM and developed working knowledge of the Foundation’s donor data environment, including MS SQL Server, SSRS, and SSIS.

• Modified existing SQL queries and SSRS reports to add filters showing each fundraiser’s quarterly/campaign results against established fundraising goals, giving the VP of Development visibility into performance by individual and fundraising unit
• Validated report outputs and underlying data to ensure accurate, reliable results and delivered the reporting solution ahead of deadline, providing actionable intelligence to prioritize fundraising follow-up and help drive progress toward organizational goals
• Initiated Python-based RFM donor segmentation to analyze donor recency, frequency, and monetary value, establishing an early framework for more targeted donor analysis
• Tech: Blackbaud CRM, MS SQL Server, SSRS, SSIS, Python

Data Scientist

Semanticbits

Member of remote, Agile data team supporting Analytics and Reporting for leadership of CMS’s Quality Payment Program (QPP) by delivering diverse data products using SQL, Python, JupyterLab, and Superset.

• Developed and validated data tables for public-facing reports
• Built dashboards with interactive visualization of KPIs for senior management
• Transitioned management dashboards from PostgreSQL to Redshift across all tiers, reducing average page load time to well below 2 second target, a >70% decrease
• Designed and maintained data mart objects supporting QPP products, e.g. Management Dashboards, Clinician Data Lookup Tool, Policy Simulator
• Validated accuracy of data and reports, including implementation of automated data checks and reviews of production SQL and Python code
• Presented sprint deliverables to program leadership, and gathered feedback for bug fixes and new features
• Trained external teams and potential customers on data model, metric definitions, and use of platform tools for self-service analytics
• Member of hiring team for over 1 year: conducted dozens of phone screens and technical interviews, plus I created a technical interview for a BI role
• Tech: Python, SQL, PostgreSQL, Redshift, Jupyter Notebooks, Superset, Git, GitHub

Data and Analytics Specialist/Software Engineer

Tripwire

Developed security analytics, anomaly detection, and visualization solutions using Python and the Elastic Stack as Tripwire's first data science hire, transforming raw security event data into actionable product insights.

• Compelling visualizations of filtered and/or aggregated event data along side insightful analytics (matplotlib)
• Over two dozen security data features engineered with standard (e.g. FFT) and new (e.g. Block analytics) methods
• An anomaly detector that successfully flagged “unusual” security events
• A movie simulation of flagging anomalies in real-time data streams as POC for the marketing and product teams
• High-level presentations and demos of visualization and analytics results, including at R&D All-Hands
• Tech: Python, Elasticsearch, Logstash, Kibana, NumPy, Matplotlib

MATLAB developer

Intel, contracted through EdgeLink

Developer tasked with transitioning fab cost models from EXCEL VBA to MATLAB while on temporary leave from my PhD work at MIT.

• Developed and implemented a method for documenting Excel VBA cost model requirements
• Coded efficient, accurate MATLAB cost model components from requirements
• Trained team members on documentation method and MATLAB fundamentals

Graduate Researcher

Massachusetts Institute of Technology

As an academic researcher on the front line of discovery about the way the Earth works, I coded data analysis algorithms on a daily basis (mostly Matlab), delivered dozens of presentations, and collaborated on numerous research projects with colleagues from around the world. Check out the publications tab of my website for additional details.

• Produced the first high-resolution seismic image of continental lithosphere subducting into the mantle beneath Northern Greece, a controversial result (see Pearce JGR 2012 paper)
• Developed novel approaches for using seismic images to quantitatively assess the validity of tectonic/geodynamic models, e.g. measuring ~70 km of differential slab rollback induced offset across the Cephalonia Transform Fault in Western Greece, consistent with geologic and geodynamic modeling estimates
• Identified a reflector interpreted as the base of the external Hellenides thin-skinned thrust belt and discovered that it breaches the surface where basement rocks are exposed within the footwall of active extensional faults, as predicted by geodynamic modeling, compelling evidence for how continental crust forms in retreating subduction settings
• Developed a preprocessing pipeline to generate receiver functions from array data with unprecedented fidelity and accuracy, while expanding the number of earthquakes available for imaging (i.e. it made lower signal-to-noise EQs viable for imaging)
• Derived analytical expressions for receiver function imaging of 3D, dipping interfaces with multiple layers to power a Common Conversion Point imaging algorithm
• Managed and quality controlled over a dozen big (TB+) data sets, 6+ for publication
• Developed 12+ data processing workflows combining signal processing (FFT, PCA, adaptive filtering) and machine learning (linear regression, genetic algorithms) — early applications of quantitative ML methods to geophysical inverse problems
• Trained research teams to apply seismic imaging workflows on their own datasets, resulting in several peer-reviewed publications
• Developed an automated multichannel alignment pipeline (Appendix B of PhD thesis) replacing a manual qualitative process with an iterative, convergence-based automated system — a direct precursor to modern MLOps thinking
• Presented over twenty times at conferences and universities, including invited talks for the Tectonics Observatory Brownbag Seminar at Caltech in July of 2011 and the Department Seminar at the University of Oregon in January of 2011
• Tech: MATLAB, Python, Signal Processing (FFT, PCA, Filtering), Linear Regression, Genetic Algorithms, Distributed Computing, Seismic Imaging

Post MS Researcher

Los Alamos National Laboratory

Performed field experiments and data analysis to measure soil hysteresis effects on local resonance modes induced by a large vibrator truck, with application to designing strong ground motion resistant structures.

• Developed and coded novel signal processing methods for extracting quantitative measures of nonlinear soil behavior from large (>GB), experimental data sets
• Wrote publication (Johnson et al., JGR 2009) and presented research findings at 6+ meetings and conferences
• Tech: MATLAB, Signal Processing, Large-scale data analysis (>GB datasets)

AgentFluent

Local-first Python CLI on PyPI that analyzes Claude Code and Claude Agent SDK session traces to diagnose agent configuration quality and recommend improvements.

• Created deterministic evaluation pipelines that correlate prompt design, tool selection, model configuration, retry behavior, token usage, and execution errors with actionable optimization recommendations
• Scores every recommendation on one of three diagnostic axes — cost (token spend, model selection trade-offs), speed, and quality (mid-flight user corrections, file rework density, reviewer-caught rate) — exposing per-recommendation axis_scores and primary_axis in the JSON envelope so a CI rule can gate on a single dimension
• Enhanced quality diagnostics with opt-in git and GitHub enrichment layers that correlate session behavior to downstream CI failure rates, PR review density, and feat-to-fix commit proximity
• Implemented rolling baseline regression detection, structured JSON outputs, CI/CD integration, automated versioning, and quality gates using GitHub Actions, mypy, pytest, ruff, and release-please
• Designed a privacy-first, zero-cost architecture that runs locally with no outbound network calls unless GitHub enrichment is explicitly opted into, enabling enterprise-safe agent evaluation, monitoring, and continuous improvement
• Extracted the supervised dev-loop engine powering AgentFluent's own development into a standalone Claude Code marketplace plugin, making the same structured, human-gated development loop available to any project
• Tech: Claude Code, Claude Agent SDK, Python, NLP, Clustering, Pydantic, PyYAML, pytest, mypy, ruff, GitHub Actions, Loop Engineering

CodeFluent

VS Code extension and FastAPI web application that analyzes Claude Code session data to measure human-AI collaboration quality against published Anthropic research benchmarks.

• Designed LLM-based behavioral evaluation pipelines that score 11 AI fluency behaviors, 6 coding interaction patterns, and project configuration maturity while detecting policy enforcement gaps within CLAUDE.md workflows
• Implemented personalized coaching that generates optimized prompts, identifies missing collaboration behaviors, and correlates fluency scores with token consumption, cache utilization, and development cost
• Architected cross-platform Python and TypeScript services with persistent caching, structured JSON parsing, GitHub CLI integration, and comprehensive automated testing exceeding 1,650 unit and integration tests
• Built a production-ready analytics platform that turns raw Claude Code conversations into actionable engineering insights through interactive visualizations, research-backed scoring models, and repository-specific recommendations
• Tech: Python, FastAPI, VS Code Extension API, Claude API, Chart.js, pytest, TypeScript

Sportswear ESG News Classifier

Fully autonomous production ML pipeline monitoring ESG news for 50+ global sportswear brands, publishing a live Sustainability Scorecard and interactive dashboard.

• Architected and deployed an autonomous production ML pipeline that monitors ESG news for 50+ global sportswear brands, classifies articles through a hybrid ML and LLM workflow, and publishes a live Sustainability Scorecard with evidence-backed rankings, sentiment analysis, and searchable dashboards
• Built a two-stage classification pipeline using a Random Forest (F2: 0.974) pre-filter ahead of LLM-as-Judge ESG classification, reducing inference costs by 10% while preserving high recall to capture all relevant content
• Created a RAG pipeline with OpenAI embeddings, PostgreSQL, and pgvector to retrieve supporting article passages for every ESG classification, providing traceable evidence instead of opaque model outputs
• Automated end-to-end operations with Python orchestration for news collection, drift detection, model evaluation, website publishing, and health reporting, replacing 5-10 hours of weekly manual effort at an operating cost of about $0.15/day
• Deployed and maintained production infrastructure using FastAPI, Docker, Google Cloud Run, MLflow model registry, Evidently AI, GitHub Actions CI/CD, and a 900+ test suite supporting continuous delivery and monitored model performance
• Tech: Python, Claude API, OpenAI Embeddings, scikit-learn, Jupyter, FastAPI, Docker, Google Cloud Run, PostgreSQL, pgvector, MLflow, Evidently AI, GitHub Actions

E-Commerce Fraud Detection

Production-grade fraud detection REST API combining feature engineering, model training, explainability, automated testing, and cloud-native deployment (ML Zoomcamp capstone).

• Built a production-grade fraud detection service combining feature engineering, model training, explainability, automated testing, and cloud-native REST APIs to evaluate real-time payment transactions with configurable business risk thresholds
• Developed an XGBoost classifier achieving 0.87 PR-AUC with 73% precision at 83% recall on a 299K transaction dataset with a 44:1 class imbalance, transforming 15 raw fields into 30 domain-specific behavioral and transaction features
• Implemented five configurable decision threshold strategies and TreeSHAP-based prediction explanations, enabling fraud teams to balance false positives against fraud losses while providing transparent, analyst-ready decision rationale
• Designed and deployed a FastAPI service with Pydantic validation, multi-stage Docker containers, and Google Cloud Run auto-scaling, sustaining 25 requests/second at under 70ms server-side P99 latency under Locust load testing
• Validated production reliability through a 448 test suite covering preprocessing, feature engineering, model inference, explainability, API endpoints, and integration workflows
• Tech: Python, XGBoost, scikit-learn, Jupyter, FastAPI, Docker, Google Cloud Run, SHAP, Pydantic, Pytest, Locust

US Presidential Election Analysis

Multi-source ETL pipeline and public REST API serving U.S. Electoral College and popular vote data from 1824 to present, live at api.us-presidential-election-center.org.

• Built a multi-source ETL pipeline ingesting U.S. Electoral College and popular vote data from the National Archives (web scraping), MIT Election Lab (CC0 dataset), and UCSB American Presidency Project, reconciling sources onto shared candidate and state keys in a PostgreSQL star schema warehouse spanning elections from 1824 to present
• Designed a read-only SQLite snapshot architecture that fully decouples the API serve path from the live database — the Docker container serves requests with Postgres stopped, eliminating runtime database dependencies and enabling scale-to-zero Cloud Run deployments within the free tier
• Deployed a public FastAPI REST API (api.us-presidential-election-center.org) fronted by Cloudflare for CDN caching, rate limiting, and bot protection; implemented keyless CI/CD via Workload Identity Federation GitHub Actions with version-based cache invalidation on cutover
• Implemented rigorous data licensing compliance: MIT Election Lab data served under CC0 1.0, non-redistributable UCSB data excluded from the public API, with provenance metadata embedded in every response envelope
• Publishes the 'Counted, Not Assumed' blog series documenting two centuries of presidential election data from primary sources, supported by an automated cross-repo publishing pipeline with fail-closed OG card validation and validate-all-then-write atomicity guarantees
• Tech: Python, FastAPI, PostgreSQL, SQLite, Docker, Google Cloud Run, Cloudflare, GitHub Actions, BeautifulSoup, pandas, geopandas, uv

PhD

Massachusetts Institute of Technology

Geophysics

MS

Massachusetts Institute of Technology

Geosystems

BS

University of Idaho

Geological Engineering

Student Presentation Award

Seismological Society of America

Received the Seismological Society of America Student Presentation Award as recognition for an outstanding research presentation at the 2010 SSA Annual Meeting. The slides for the presentation entitled "Imaging with scattered teleseismic waves: data, method, and application to the Hellenic subduction zone" are available on my LinkedIn profile.

Geochemistry Achievement Award

Professor Scott Wood, University of Idaho

Received 1st annual award for outstanding achievement in geochemistry course.

AI Engineer (Volunteer)

DataTalksClub — Machine Learning Zoomcamp

Built a RAG-based GitHub automation system to triage FAQ submissions for DataTalksClub's open-source ML course community.

• Designed an agentic workflow using RAG to classify incoming GitHub issues as NEW, UPDATE, or DUPLICATE, auto-generating formatted pull requests or closing duplicates with contextual comments
• Implemented a production-ready system with Pydantic-structured LLM outputs, 153 tests, and autonomous deployment via GitHub Actions, eliminating the manual PR review bottleneck for course maintainers
• Tech: Python, OpenAI API, RAG, GitHub Actions, Pydantic, pytest

Scrum Master

AgilePDX

Co-organized AgilePDX's AI & Agility Speaker Series, a new program exploring how AI enhances team performance, streamlines processes, and improves product outcomes. Responsible for speaker identification, prioritization, outreach communications, and presentation logistics across a four-month run.

Technical Evaluator

Emerging Leaders PDX

Volunteer technical interviewer asking tech questions and providing constructive feedback during mock interviews of Emerging Leader candidates — junior and senior college students from underrepresented groups — seeking to receive internships in the tech industry, as part of a broader mission to increase diversity, equity and inclusion within the leadership of Portland's tech community.

Data Scientist, Subject Matter Expert

Hack Oregon

Team member on the Disaster Resilience project, focused on delivering a web-based app to help members of the Portland community 1) understand the risk they face during a Cascadia subduction zone earthquake, and 2) access resources to prepare themselves and their neighbors. Responsible for building an application to provide ground shaking intensity estimates for individual homes or businesses from raster-based earthquake model outputs, using Docker, Python, GDAL, Rasterio, Geopandas, PostGIS, etc.

• The team's Earthquake Resiliency app was featured at the Oregon Museum of Science and Industry.

Presenter

Portland Machine Learning Meetup

Gave a presentation on my PhD thesis work entitled "Data Science View of Subduction Zone Imaging", including a description of the different datasets that are melded together, the preprocessing/imaging methods, and the interpretation of the resulting seismic images in terms of subduction zone structure.

Session Convener and Chair

American Geophysical Union

Convenor and Chair for session entitled "Advances in Signal Processing Methods for Seismology". Solicited abstracts for session, organized them into a logical program, and introduced speakers when it was show time.

Session Chair

American Geophysical Union

Chair for session entitled "The Formation and Deformation of the Mediterranean Basins, Continental Margins, and Arcs II".

Session Chair

American Geophysical Union

Chair for session entitled "Nonlinear Geophysics General Contributions Posters"

Organizer

Earth Resource Laboratory, MIT

Co-organizer for the Earth Resources Laboratory's Friday Informal Seminar Hour (FISH), a weekly technical presentation on a wide variety of geophysical research topics. I helped set up speakers, arrange their schedules, and provided introductions prior to the presentation.