Principal AI Engineer at Verizon
Open to Director, Head of AI, and Principal roles

Scott Severance

Principal AI Engineer
Applied AI and Agentic Systems

I build the production ML and agentic AI systems a Fortune 15 People Analytics organization runs on, then prove what they were worth: $10.4M+ in validated program impact, a 47-function ML engine inside the Qlik query layer, and three Model Context Protocol servers wiring LLMs into live enterprise data. Nine years of production ML on two decades of enterprise analytics.

Get in Touch See the Numbers
Currently at Verizon
Education M.S. Data Science, SMU
Recognition Brandon Hall Gold Award
Open Source 6 AI Frameworks
About

From First Model
to GenAI Platform at Scale

I built Verizon HR's first production machine learning model in 2017. Today I build the infrastructure underneath enterprise AI (gRPC services, MCP servers, agent pipelines) and own the outcome metric from prototype to production.

As Principal AI Engineer at Verizon, I work as an internal forward-deployed engineer across 12 business domains, and I am the sole technical owner on most of what I ship. That means I write the production code as well as the architecture: a Python/gRPC engine exposing 47 machine learning functions as first-class Qlik expressions, deployed as a Windows service on a cluster I administer; three Model Context Protocol servers, the largest exposing 111 tools across a Claude/OpenAI/Azure abstraction with OAuth 2.1; and an NLP classifier behind a Workday triage agent.

The other half is measurement. I hold graduate-level applied statistics and use it for real causal inference (synthetic control, difference-in-differences, ANOVA with post-hoc correction, every sample Cochran-sized), which is how $10.4M+ of program impact was validated rather than asserted. I also fine-tune models myself, LoRA and full-parameter, on an 8×A100 DeepSpeed cluster.

Rated Exceeding or Top Performing three years running, with a Brandon Hall Gold Award for AI Innovation and six open-source AI frameworks on GitHub.

GenAI Pioneer

Architected Verizon's first enterprise GenAI deployment in HR on Google Vertex AI and Gemini

Top Performer

Top Performing rating in 2025, the third consecutive year rated Exceeding or Leading

Team Builder

Cut model deployment from 8 weeks to 3 while mentoring the next generation of data scientists

Impact

Measurable Results

Production systems shipped, real metrics, and business value quantified at Fortune 15 scale.

$10.4M+
Validated Business Impact

Quantified across 77 research initiatives using the method that fits each design (synthetic control, difference-in-differences, ANOVA with post-hoc correction), every sample Cochran-sized for a controlled margin of error. Includes a corrected methodology that produced a 1,615% ROI finding, and a Tuition Assistance ROI baseline I re-validated nine years later.

47
ML Functions in the Qlik Engine

A Python/gRPC Server-Side Extension turning clustering, forecasting, NER, and the full scikit-learn lifecycle into first-class query expressions, evaluated live against the user's current selection.

111
Tools Exposed via MCP

The largest of three Model Context Protocol servers, spanning a Claude/OpenAI/Azure provider abstraction with OAuth 2.1: hypercube queries, master-item CRUD, visualization creation, selection state.

70%
Reduction in Query Time

Verizon's first enterprise GenAI deployment in HR delivered natural-language analytics over 50+ datasets to 1,000+ users.

54% → 89%
NLP Classification Accuracy

A 35-point lift from a debiasing pipeline (spaCy + Hugging Face, human-in-the-loop labeling); sentiment moved from 63% to 88% on the same corpus. Findings went to the Board of Directors.

95%+
Forecast Accuracy

Voluntary turnover model (PyTorch + Prophet) wired into the Workforce Management Dashboard: $500K+ in annual recruiter savings.

10M+
HR Documents Analyzed

RAG summarization agent enabling instant policy retrieval across the enterprise HR document library.

8 → 3
Weeks to Deploy a Model

MLOps governance built from scratch on GitLab CI/CD and Domino: bias detection, drift monitoring, deployment standards.

~40%
Estimated Compliance Risk Reduction

Risk clustering engine (XGBoost + SHAP) auto-flags Code of Conduct violations for early intervention, avoiding roughly $250K a year in external firm costs.

50%+
Coding Velocity Boost

An AI coding assistant (LangChain + Gemini) shipped to the analytics team with usage and quality instrumentation.

21 / 120
Contributors & Commits Governed

One of two Maintainers on a shared agentic-AI repo. Authored the governance framework everyone worked inside and personally onboarded ~20 non-technical analysts onto a full branch, merge-request, and review workflow.

Signature Systems

Four Things I Built
End to End

Architected, coded, deployed, and operated, with the problem each one solved and the result it produced. These are the systems I would walk you through in a technical interview.

01

ML Inference Inside the Query Layer

Qlik's expression language has no machine learning surface. Every model-backed metric had to be precomputed outside the tool and re-imported, which was slow and cut off from the user's live filter context.

I built a Python/gRPC Server-Side Extension exposing 47 native functions (HDBSCAN clustering, Prophet forecasting, the full scikit-learn train/predict/score lifecycle, spaCy NER, market-basket analysis, Keras inference), callable as ordinary chart expressions and evaluated against the current selection. Deployed as a production Windows service on an on-prem cluster I administer.

Result: Qlik went from a static BI layer to a live ML inference surface for the whole People Analytics organization; any app can call a trained model against the user's live selection.

PythongRPCscikit-learnKerasHDBSCANProphetspaCy
02

Governed LLM Access to Live Enterprise Data

Claude and ChatGPT both needed a safe, auditable path to read and act on live Qlik data, cloud and on-prem, without hand-rolling a fresh integration per provider.

Three independent Model Context Protocol servers. The largest exposes 111 tools behind a Claude/OpenAI/Azure provider abstraction with OAuth 2.1: app discovery, field inspection, hypercube queries, master-item CRUD, sheet and visualization creation, selection-state management. A second ships to the team with bundled certs and SSO/EID auth against on-prem Qlik Sense.

Result: any MCP-compatible client gets governed, tool-level access to both Qlik estates, the connective layer under most of the 2024-2026 AI work.

Model Context ProtocolOAuth 2.1Multi-provider LLMQlik Engine JSON API
03

Agentic HR Case Triage

Workday generates a large backlog of unstructured HR cases that need manual categorization and routing before a specialist can act on any of them.

I designed the agent architecture end to end, benchmarking Claude Agent SDK against Google Vertex AI ADK on cost, tool-calling reliability, and enterprise auth fit, then built the NLP classifier underneath it on historical case data to auto-categorize and route incoming tickets.

Claude Agent SDKVertex AI ADKNLP ClassificationWorkday
04

Getting 20 Analysts Into a Real Git Workflow

An org-wide agentic-AI push meant handing modern tooling to people whose job title is not engineer; most had never opened a merge request.

As one of two Maintainers on a 21-contributor, 120-commit repo, I authored the governance framework the whole team worked inside: a root CLAUDE.md, path-scoped rule files for data quality, HR privacy, and Qlik standards, plus an on-demand skills library. Then I personally onboarded ~20 non-technical analysts onto branch, merge-request, and review.

Result: they shipped working artifacts through real code review inside two weeks.

Git/GitLab GovernanceClaude CodeTechnical Enablement
Expertise

Core Capabilities

Full-stack AI engineering, from LLM evaluation harnesses and agentic orchestration to production deployment and enterprise governance.

Agentic AI & LLM Engineering

MCP server design (111- and 46-tool surfaces), agent architecture benchmarked across Claude Agent SDK and Vertex AI ADK, multi-provider abstraction with OAuth 2.1, RAG pipelines, and evaluation harnesses. Fine-tunes models directly, LoRA and full-parameter, on 8×A100 with DeepSpeed.

Claude Agent SDKMCPVertex AI ADKGeminiLangChainLangGraphRAGLoRADeepSpeedEvals

Production Machine Learning

Full lifecycle ownership, from feature engineering through API deployment and drift monitoring. Forecasting at 95%+ accuracy, ensemble attrition models, risk clustering with SHAP explainability, and hyperparameter work that lifted a live search classifier from 75% to 86%.

PyTorchTensorFlowscikit-learnXGBoostLightGBMProphetSHAPOptuna

NLP & Deep Learning

A model lineage that runs from lexicon scoring to classical ML to LSTM/GRU/Bi-GRU with GloVe embeddings to transformer fine-tuning. BERT query classifier at ~91% validation accuracy; debiasing pipelines that moved classification accuracy 35 points on 60,000+ survey responses.

BERTLSTM/GRUGloVespaCyHugging FaceVADERBERTopic

Causal Inference & Measurement

Graduate-level applied statistics used on real programs: synthetic control, difference-in-differences regression, ANOVA with Tukey HSD and Dunn's post-hoc correction, every sample Cochran-sized for a controlled margin of error.

Synthetic ControlDiff-in-DiffANOVA / Post-HocCochran SamplingA/B TestingKirkpatrick L4 ROI

Data & Platform Engineering

Python/gRPC services, REST and OAuth integrations, API/sFTP pipelines, and Python ETL, plus on-prem Qlik Sense Enterprise cluster administration and a lead role in the PeopleSoft to Workday HCM cutover, rebuilding and revalidating every downstream BI feed.

gRPCREST/OAuthQlik SenseQlik AutomationspandasTableauLookerSQL

MLOps & AI Governance

End-to-end model lifecycle management with GitLab CI/CD, Domino, and Airflow. Bias detection, drift monitoring, and ethical AI frameworks.

GitLab CI/CDDominoDockerAI Ethics

Cloud & Infrastructure

Multi-cloud platform deployment across GCP, AWS, and Azure. Container orchestration, scalable data pipelines, and enterprise-grade infrastructure for production AI workloads.

GCPAWSAzureKubernetes
Projects

GitHub Projects

Open-source tools and experiments spanning MLOps, agentic AI, edge computing, and security.

Self-Healing-MLOps-Pipeline

Production-grade MLOps pipeline with automated drift detection, self-healing retraining, and feature store integration.

Python
scikit-learnFeastGitHub Actions

DeepResearchAgent

Autonomous research agent that plans, searches, synthesizes, and produces cited reports using LangGraph.

Python
LangGraphAgentic AI

Edge-Native-Vision-Copilot

Real-time computer vision assistant optimized for edge deployment with ONNX runtime and voice interaction.

Python
ONNXEdge AIComputer Vision

GraphRAG-Compliance-Navigator

Knowledge-graph-powered RAG system for searching complex regulatory and compliance documents.

Python
Neo4jGraphRAGLangChain

Autonomous-SOC-Security-Agent

AI-driven security operations center agent with automated threat triage, enrichment, and human-in-the-loop escalation.

Python
SIEMThreat IntelHITL

Data-Science-Toolkit

Streamlit app for automated EDA, feature engineering, and model benchmarking with SHAP explainability and Optuna search; runs fully local with no telemetry.

Python
StreamlitSHAPOptuna

ToneAnalyzer

Real-time speech and text emotion analysis using TensorFlow.js with a React-based interactive dashboard.

TypeScript
ReactTensorFlow.jsEmotion AI

Vulnerability-Scanner

Five-phase security pipeline (scan, penetration test, zero-day detection, hardening, AI analysis) with 14 CIS-aligned scanners, backup-first hardening with rollback, and a standalone HTML report. Two runtime dependencies.

Python
CIS HardeningZero-Day DetectionLLM Analysis

CodingAgent

Offline-first AI coding assistant powered by Ollama and local LLMs for private, air-gapped development.

Python
OllamaOfflineLocal LLM

Agent-Builder

Cross-platform CLI framework to design, build, and deploy AI agents with 6 LLM providers, 5 agent patterns, DAG workflows, and MCP compatibility.

TypeScript
MCPDAG WorkflowsMulti-LLM

PDA

AI phone assistant. Twilio forwards the call, Claude (or Grok or Gemini) holds the conversation, and a Telegram summary lands on your phone when it ends.

Python
TwilioClaudeTelegram
View GitHub Profile
Career

Professional Experience

Principal AI Engineer

Embedded AI Engineer and Internal Forward-Deployed Engineer across 12 Business Domains
Verizon | Irving, TX
2019 - Present
  • Qlik Server-Side Extension: a Python/gRPC engine exposing 47 native ML and NLP functions (HDBSCAN, Prophet, full scikit-learn lifecycle, spaCy NER, market-basket, Keras) as first-class query expressions, evaluated live against the current selection; deployed as a production Windows service on a cluster I administer
  • Three Model Context Protocol servers linking Claude and ChatGPT to Qlik Cloud and on-prem Qlik Sense, the largest exposing 111 tools across a Claude/OpenAI/Azure provider abstraction with OAuth 2.1; a second shipping 46 tools with bundled certs and SSO/EID auth
  • Workday HR case-triage agent: designed end to end, benchmarking Claude Agent SDK against Google Vertex AI ADK on cost, tool-calling reliability, and enterprise auth fit; built the supporting NLP classifier on historical case data
  • Built Verizon's first enterprise GenAI deployment in HR (Vertex AI + Gemini): conversational analytics over 50+ datasets, with analyst query time down 70% across 1,000+ users
  • Maintainer on a 21-contributor, 120-commit AI-tooling repo: authored the governance framework (CLAUDE.md, path-scoped rule files, skills library) and onboarded ~20 non-technical analysts onto a full branch, merge-request, and review workflow
  • Multi-year NLP toolkit built in stages, from a VADER baseline to SVM to Keras LSTM/GRU/Bi-GRU with GloVe embeddings; separately fine-tuned BERT for a query classifier at ~91% validation accuracy
  • NLP debiasing pipeline (spaCy + Hugging Face, human-in-the-loop labeling): topic accuracy from 54% to 89% (+35 pts) and sentiment from 63% to 88% (+25 pts), presented to the Board of Directors
  • Turnover Forecasting Model (PyTorch + Prophet) at 95%+ accuracy, from feature engineering through API integration into the Workforce Management Dashboard: $500K+ in annual recruiter savings
  • AI Risk Clustering Engine (XGBoost + SHAP) auto-flags Code of Conduct violations for early intervention: 40% estimated compliance risk reduction and roughly $250K avoided annually in external firm costs
  • Led multi-year causal inference evaluations matching method to design (synthetic control, difference-in-differences, ANOVA with Tukey HSD and Dunn's post-hoc), validating $10.4M+ across 77 initiatives, including a corrected methodology that produced a 1,615% ROI finding
  • Integrated Gemini into production Qlik Cloud apps via Qlik Automations (in-app natural-language chat that respects live selection state); built a Playwright Slack scraper feeding a Claude-API summarization pipeline for automated status briefs
  • Lead role in the 2023 PeopleSoft to Workday HCM migration: rebuilt and revalidated every HR BI feed and product against the new source through cutover
  • MLOps governance stack (GitLab CI/CD + Domino) built from scratch with bias detection, drift monitoring, and deployment standards; deployment cycle compressed from 8 weeks to 3 weeks. As Data Science Academy Cohort Lead, mentored 3 junior data scientists
Top Performing Rating (2025) Exceeding Rating (2024) Leading Rating (2023) Brandon Hall Gold Award

Data Scientist

ML and Analytics Pioneer
TEKsystems (at Verizon) | Irving, TX
2017 - 2019
  • Brandon Hall Gold Award for AI Innovation: architected and deployed Verizon HR's first production ML model, a TensorFlow neural network classifying open-ended learning-feedback survey text at 85% accuracy
  • Built the business case behind a $3M training-investment decision from travel spend, vendor usage, and tuition data: $10M in annual savings
  • Owned the enterprise L&D executive dashboard program, reporting on $160M+ in annual training investment and 7.9 million training hours; built and maintained 60+ Qlik dashboards and Tableau workbooks with Python ETL pipelines
  • Directed a cross-functional, 8-business-unit survey-verbatim labelling program: a 3-level topic/sentiment taxonomy that trained a classifier validated against 60,000+ quarterly responses
  • Designed an unbiased-control framework for HR program ROI studies, strengthening the statistical rigor behind 10+ high-stakes impact analyses
  • Mentored junior data scientists in NLP and advanced machine learning techniques

Head of Business Intelligence

Rs Reps | Grand Prairie, TX
2015 - 2017
  • Built the data and business intelligence function and grew the company's digital presence alongside it
  • Analyzed market research and manufacturer data in Tableau, Python, and SAS to support data-driven decisions
  • Deployed multivariate linear regression models (SAS, R) for sales forecasting, informing strategic planning and resource allocation

Data Analytics Manager

Verizon | Irving, TX
2011 - 2015
  • Ran cause-and-effect efficiency analyses on major marketing campaigns and website personalization redesigns, tripling online sales close rates and doubling the share of FiOS broadband sales completed through e-commerce channels
  • Built statistical models and dashboards for Verizon's e-commerce sales pipeline and presented the findings to stakeholders
  • Developed e-commerce channel economics and introduced new operating ratios to measure growth accurately
  • Produced sales and site-engagement forecasts through regression analyses, consistently within a 10% margin of accuracy
  • Formed and tested hypotheses on chat engagement, self-service success, and website updates to guide optimization

Senior Ecommerce Consultant

TEKsystems (at Verizon) | Irving, TX
2007 - 2011
  • Designed and built statistical models and dashboards on Verizon's proprietary data assets to inform business decisions
  • Contributed visualizations and data-driven perspectives to the senior leadership team's operational reviews
  • Improved regression forecasts by integrating marketing spend, planned initiatives, socio-economic factors, and engineering plans
  • Delivered digital marketing recommendations with cross-functional teams using campaign management, segmentation, custom metrics, and success events
Credentials

Education & Skills

Master of Science in Data Science

Southern Methodist University (SMU)

2017 | Dallas, Texas

Bachelor of Business Administration, Management Information Systems

Dallas Baptist University

Dallas, Texas

DeepLearning.AI TensorFlow Developer

Professional Certificate | Coursera

2020

Data Science Academy

Cohort Lead and Participant | Verizon and Coursera

3 Certifications, 2020

Guest Lecturer: AI and ML Applications

George Mason University | Chief Learning Officer Certification Program

2020-2024 | 4 Cohorts

Leadership & Recognition

"If I had the chance to have him on my team again, I would quickly snatch him up before someone else did."

Nicole DiStefano, Senior Manager, on naming Scott the "Ideal Data Science Professional"
Brandon Hall Gold Award for AI Innovation
Top Performing Rating (2025)
Exceeding Performance Rating (2024)
Leading Performance Rating (2023)
Guest Lecturer: AI and ML Applications, George Mason University CLO Program (2020-2024)
COVID-19 "Work Forward" Program: architected the enterprise data response pipeline, highlighted in the Wall Street Journal
Member: Data Science Association, Digital Analytics Association, IEEE Computer Society

Languages & Data

Python R SQL NoSQL (MongoDB, Cassandra)

GenAI & LLM Engineering

Claude Agent SDK Model Context Protocol Vertex AI ADK Gemini Claude OpenAI GPT Claude Code Llama RAG Agentic Workflows LangChain LangGraph OAuth 2.1 Multi-Provider Abstraction Prompt Engineering Evals LoRA / Full-Parameter Fine-Tuning DeepSpeed

ML & AI Frameworks

PyTorch TensorFlow Keras scikit-learn XGBoost LightGBM Hugging Face Transformers BERT LSTM / GRU / Bi-GRU GloVe spaCy Prophet HDBSCAN SHAP Optuna

Platforms & Tools

Google Vertex AI AWS SageMaker Azure BigQuery ML Qlik Sense Looker Tableau Spotfire Qualtrics Workday SAS

MLOps & Infrastructure

gRPC REST / OAuth GitLab CI/CD Git Branch Governance Domino Airflow Docker Kubernetes Playwright Streamlit ADRs / C4 Diagrams API/sFTP

Statistics & Methodology

Synthetic Control Difference-in-Differences ANOVA / Tukey HSD / Dunn's Cochran Stratified Sampling Logistic Regression Predictive Modeling A/B Testing Time Series Clustering Organizational Network Analysis Kirkpatrick L4 ROI AI Ethics & Bias
Connect

Hiring for Applied AI at Enterprise Scale?

Open to Director and Head of AI roles, Principal, Staff, and Distinguished Engineer roles, and Applied AI / Forward Deployed Engineer roles.

If you need someone who builds the system, trains the model, and owns the outcome metric, let's talk. Based in Dallas-Fort Worth and open to remote.

Send an Email Connect on LinkedIn

Email

scott@scottseverance.net

LinkedIn

/in/scott-severance

Location

Dallas-Fort Worth, Texas