$350M
Procurement program powered by ML forecasting model (decision trees + logistic regression)
Selected Outcomes
Impact in production
$2.5M
Projected annual savings from the frac data acquisition platform
$40K/mo
Recurring savings from eliminating third-party vendor software
97%+
Field data reliability after pipeline + device overhaul (from <80%)
Systems I've Built
Production platforms, sole architect
The systems below are production platforms I designed and shipped end-to-end — ML models, cloud architecture, data pipelines, authentication, APIs, frontend, testing, and deployment. Each is in daily use by engineering and operations teams at Devon Energy.
Flagship · AI-Integrated · Production
Real-Time Frac Data Acquisition Platform
Sole architect of an end-to-end ingestion, validation, and AI-assisted backfill platform replacing a third-party vendor pipeline. Supports live frac fleet operations across Devon's field footprint — built on integrated agents, automated fallback, an LLM-powered backfill engine, and a React monitoring app used daily by completions engineers.
Agents
Python/PySpark ingestion agents pull ~80 high-frequency telemetry files per frac fleet daily, run gap/spike validation, and stream clean data into the OSIsoft PI historian in near-real-time.
Fallback
Primary/secondary ingestion paths with health checks, automatic failover on stream drop, exponential backoff retries, and self-healing recovery on node or device loss — keeps ingestion flowing when upstream sources degrade.
AI Backfill
LLM-based agents (OpenAI SDK) parse unstructured vendor PDF reports and heterogeneous CSV exports, extract telemetry fields, and map varied service-provider schemas into PI-compatible format for historical gap recovery.
Frontend
React + Node.js operator-facing monitoring app — live ingestion status, data-quality inspection, manual backfill triggers, and fallback/failover visibility. Used daily by completions engineers.
Infra
Partnered with IT to replace third-party cellular links with direct VPN; data reliability <80% → >97%.
Impact
$40K/month recurring savings · $2.5M/year projected acquisition-fee savings · replaced an entire vendor pipeline
ML + MLOps · Azure · $350M Program
Casing Procurement & Demand Forecast Platform
React/TypeScript + Node.js platform on Azure supporting a $350M casing procurement program. Built a production forecasting model combining decision trees and logistic regression with end-to-end MLOps — continuously re-tuning against drilling and completion schedule changes to produce exact per-SKU counts and delivery windows for the forward year.
ML Model
Decision trees + logistic regression ensemble. Re-scores casing needs by SKU (OD, weight, grade, connection) across 30–180 day and full-year horizons as drilling and completion operations shift day-to-day.
MLOps
Fully deployed on Azure with automated retraining pipelines, drift monitoring, model observability, and production-grade error handling — continuously keeps model outputs aligned with real field conditions.
Stakeholder Inputs
Drilling engineers submit per-well casing plans, supply chain teams feed in-flight inventory and buffer thresholds, and casing providers contribute lead-time and shipment commitments — all rolled into each forecast run.
Delivery Layer
Year-ahead delivery projection schedules rig-level casing deliveries across the full forward program, auto-rebalancing whenever drilling order, completion pace, or provider lead times shift.
Azure Stack
App Service, Container Registry, Azure AD SSO (MSAL + Graph API, 3-tier RBAC), Blob Storage (SAS URLs), Logic Apps + Graph API cron pipeline.
Data
15+ Snowflake tables across 4 enterprise databases, parameterized SQL/CTEs, in-memory caching with stampede protection, full audit trail (before/after diffs, user attribution) on every CRUD op.
Quality
Playwright E2E (admin, forecasting, audit, mobile) + Vitest unit tests — production-grade testing at enterprise scale.
Deep Learning · NLP · Production
BERT Document Intelligence System
Fine-tuned and deployed a BERT-based question-answering pipeline extracting structured fields from 5,000+ scanned TIFF images per day at RMS. Replaced a manual review workflow, reducing turnaround and human error while achieving an F1 score of 0.90.
Stack
Python, PyTorch, Hugging Face Transformers, OpenCV, Tesseract OCR
Scale
5,000+ images processed daily · F1 = 0.90 · full manual process automated
Role
Model fine-tuning, dataset curation, pipeline architecture, production deployment, monitoring
AI Agents · Developer Tooling
Agentic Validation & Developer Toolkit
Built AI-powered validation agents using Pydantic AI for casing pressure analysis with structured input/output models and custom tool functions. Authored an agentic coding toolkit with Snowflake integration plugins, query-validation hooks, and developer productivity automations. Integrated OpenAI SDK for intelligent data enrichment.
Stack
Python, Pydantic AI, OpenAI SDK, Snowflake, Fuse.js
Highlights
Structured agent I/O, custom tool functions, query-validation hooks, dev-loop automations
ML Coverage
Broad, not narrow
I've shipped across classical ML, deep learning, and LLM agents — production, not just notebooks. My view: LLMs are one tool in the box. The right model depends on the problem, the data, and what you can operate at scale.
Classical ML
Decision trees (random forest, gradient boosting), logistic regression, linear regression, clustering (k-means, hierarchical), time-series forecasting, scikit-learn, XGBoost
Deep Learning / NLP
BERT fine-tuning, PyTorch, Hugging Face Transformers, CNN architectures, OpenCV, Tesseract OCR
LLMs & Agents
OpenAI SDK, Pydantic AI, LLM-based agents for structured extraction (PDF/CSV → schemas), tool-using agents, prompt engineering, structured I/O
MLOps / Production
Automated retraining, drift monitoring, model observability, containerized deployment on Azure, production error handling, A/B-style model comparison
Statistics & Analysis
Regression analysis, hypothesis testing, distributional analysis, exploratory data analysis, time-series decomposition, feature engineering
Visualization
Power BI, Sisense, Recharts, Plotly.js, MUI Data Grid — dashboards and operator-facing UIs for non-technical stakeholders
Experience
Where I've shipped
- Sole architect of two production platforms used daily by drilling, completions, and supply chain teams.
- Shipped a decision-tree + logistic-regression forecasting model with end-to-end Azure MLOps (automated retraining, drift monitoring, model observability) powering a $350M casing procurement program.
- Designed a multi-stakeholder input pipeline aggregating signals from drilling engineers, supply chain, and casing providers into every forecast run — strong cross-functional communication across technical and non-technical teams.
- Built a real-time frac data acquisition platform with integrated AI agents, automated fallback, and an LLM-based backfill engine parsing unstructured vendor PDFs and CSVs into PI-compatible schemas.
- Containerized React + Node.js applications on Azure App Service (multi-stage Docker, ACR, Azure AD SSO, Blob Storage SAS URLs, Logic Apps, Microsoft Graph API).
- Delivered $2.5M/year projected savings, $40K/month recurring, field data reliability lifted from <80% to >97%.
- Fine-tuned, trained, and deployed a BERT question-answering model extracting structured fields from 5,000+ TIFF images per day (F1 = 0.90), replacing a manual review process.
- Implemented time-series forecasting for daily inbound document volume; built Power BI and Sisense dashboards surfacing operational performance to leadership.
- Owned the dataset curation, evaluation harness, and post-deployment monitoring — translating model metrics into business terms for ops stakeholders.
Skills
Technical depth
Production-grade ML, cloud, and full-stack — enterprise code with error handling, testing, and observability.
Languages & Data
Python, SQL, TypeScript, JavaScript, Bash · Snowflake, OSIsoft PI, PySpark, Pandas, Hive/Hadoop-style SQL-on-big-data patterns
ML / AI
Decision trees, logistic regression, random forest, XGBoost, clustering, BERT, PyTorch, Hugging Face, LLM agents, OpenAI SDK, Pydantic AI, scikit-learn, time-series forecasting
MLOps
Automated retraining, drift monitoring, model observability, containerized deployment, production error handling, regression testing on model outputs
Azure Cloud
App Service · Blob Storage (SAS URLs) · Container Registry · Azure AD / MSAL · Logic Apps · Microsoft Graph API
Big Data / Spark
PySpark pipelines, high-frequency ingestion (~80 files/fleet/day), gap/spike validation, stream processing into time-series historians
Frontend
React 18, TypeScript, Vite, Material-UI, Recharts, Plotly.js, React Query, React Hook Form, Zod
Backend
Node.js, Express, Flask, REST APIs, node-cron, multi-agent pipeline design
DevOps & Testing
Docker (multi-stage), Nginx, Azure Container Registry, CI/CD, Playwright (E2E), Vitest, Testing Library
Visualization
Power BI, Sisense, Recharts, Plotly.js, MUI Data Grid — dashboards for technical and non-technical audiences
Statistics
Regression, hypothesis testing, distributional analysis, exploratory analysis, feature engineering
Education
Foundation
M.S. Artificial Intelligence
In Progress · Started October 2025
Oklahoma Christian University
M.S. Data Science
Oklahoma Christian University
B.S. Computer Science
Love's Entrepreneurs Cup — 1st Place
$12,000 prize · 2020
Contact
Open to senior data science & ML engineering roles
Looking for teams solving real production problems with ML — predictive, prescriptive, or agentic. Happy to walk through system architecture, talk through specific problems you're solving, or share more detail on any of the work above.