AI Engineer · Educator · M.S. Applied AI, Neumont University

Harrison Smith builds AI systems that ship

Ten-plus years of software engineering, now sharpened by a Master's in Applied AI — RAG architecture, multi-model LLM orchestration, and ML data pipelines built end to end, from raw data to production. Currently building Lorekeeper.quest as a graduate capstone.

View Lorekeeper.quest Download Résumé Get in Touch

10+ yrs: combined technical experience
M.S.: Applied AI · exp. Sept 2026
4: AI products built end to end

About

From production software to production AI

I'm an AI-focused engineer based in Ogden, Utah, working daily at the intersection of model behavior and deployable software. My background is classic software engineering — backend services, SQL systems, CI/CD pipelines — and for the past several years I've been deliberately rebuilding that foundation around AI: RAG pipelines, multi-model orchestration across Claude, Gemini, and ChatGPT, local LLM hosting, and the ML data work that makes any of it trustworthy.

My M.S. in Applied Artificial Intelligence at Neumont University (expected September 2026) has been the forcing function to go deep on the parts of the stack a bootcamp skips: full ETL pipeline design, data sanitization, model training and evaluation in TensorFlow and PyTorch, and the statistics behind knowing whether a model is actually working. I pair that coursework with real production responsibility — building the evaluation framework for a live AI agent at Blytz, and shipping data pipelines that executives actually read.

I build with AI-assisted tooling as a first-class part of my workflow — Claude Code in particular — not to skip the engineering, but to move faster through it while staying accountable for every architectural decision underneath.

2026

M.S., Applied Artificial Intelligence

Neumont University, Salt Lake City, UT — expected September 2026

Model training & evaluation (TensorFlow, PyTorch, ML.NET) · RAG architecture · full ETL & data-sanitization coursework

2022

B.S., Computer Science

Weber State University, Ogden, UT

Graduate-level data science coursework completed as an undergraduate: data cleaning, predictive modeling, mathematical frameworks

Graduate Capstone

Building Lorekeeper.quest

Lorekeeper is my M.S. capstone — an open-source, self-hosted AI campaign companion for tabletop RPGs, built end to end: a LoRA fine-tune of Mistral 7B trained on session-journal data across 12 game systems, hybrid retrieval-augmented generation over the user's own campaign records, and an evaluation harness that measures both the model and the retrieval layer. No external AI APIs — the whole stack runs on the user's hardware.

Explore the live project Source on GitHub Model on Hugging Face Live & open source (AGPL-3.0)

Training Data

Built the session-journal training corpus across 12 tabletop game systems — fantasy, sci-fi, horror — with cleaning and formatting pipelines that made fine-tuning trustworthy.

Fine-Tuning

LoRA fine-tune of Mistral 7B Instruct, merged and quantized to Q4 GGUF — a 4 GB model that runs on consumer hardware, published openly on Hugging Face.

Hybrid RAG

Semantic vector search (pgvector) fused with full-text keyword search via Reciprocal Rank Fusion — embeddings handle paraphrase, keywords rescue the rare proper nouns they blur.

Measured Relevance

The retrieval gate's threshold is set by a labeled evaluation sweep, not guessed — and when nothing relevant exists, the app labels its own reply as ungrounded instead of inventing.

Evaluation Harness

In-app model comparison (fine-tuned vs. base) with semantic-similarity scoring, plus a standalone retrieval benchmark — quality is measured at both layers, continuously.

Production & Transparency

Dockerized full stack (FastAPI, React, Postgres, Ollama) with token streaming and a per-answer retrieval trace — every reply shows which records it drew from and why.

What the full pipeline actually teaches you

Every stage above is easy in isolation and hard in combination. The real lesson of Lorekeeper.quest has been that RAG quality is a data quality problem — orchestration logic can't compensate for a bad chunking strategy, and a beautifully trained model is worthless if the visualization layer buries the answer a user actually needs. Building it end-to-end, with Claude Code as an AI pair-programmer throughout, also reshaped how I write software: I spend less time on boilerplate and more time on architecture decisions, prompt and retrieval design, and evaluating whether the system's output is actually correct — the skills that matter most once AI is writing alongside you.

Selected Work

AI systems built end to end

Ogden — Multi-Agent AI Team Platform

A self-hosted, Dockerized platform for assembling AutoGen-backed AI agent personas — each with its own prompt, model, and role — into per-project teams that carry a real backlog and work it autonomously or on demand. Built as an applied research project during my M.S. coursework to explore multi-agent orchestration at a production-engineering level: sandboxed git/test tool execution, per-project cost tracking and budget caps, and full run-history observability, not just a chat demo.

ogden.harrisonsmith.ai ↗

Hermit — Local-AI Note Companion In development

A native note-taking companion for iOS, iPadOS, macOS, and watchOS with an on-device AI layer — capture, organize, and converse with your notes without your thoughts ever leaving your hardware. The same local-first AI conviction behind Lorekeeper, brought to the Apple ecosystem: private by architecture, not by policy.

ML Classification Pipeline

Full-lifecycle model development — data sanitization, ETL, training, evaluation — in Jupyter and ML.NET, deployed as a working image-classification application.

CI/CD Deployment Pipeline — ThunderBeast.Studio

GitHub Actions pipelines that compile and deploy production builds directly to a live website, owning the full deployment path from commit to production.

Experience

Where it's been applied

Business Intelligence Analyst

Nov 2025 – Present

Blytz

Deliver BigQuery-backed data pipelines and dashboards translating raw enterprise data into decision-ready outputs for executives and cross-functional teams.
Integrate enterprise data through ETL pipelines from external partners, including HubSpot, to centralize reporting and analytics.
Ensure the accuracy of Betzy, Blytz's production AI agent, by validating triggered user responses, correcting queries, and providing human-in-the-loop oversight.
Collaborate with the internal AI team to troubleshoot issues and launch new capabilities, such as empowering users to build custom dashboards from training and payment processors.
Accelerate the development of complex queries and dashboards using AI coding tools like Claude Code, and execute deep-dive data analysis using Python and Jupyter Notebooks.

Software & IT Services

Mar 2024 – Present

Common Thread

Build and maintain backend services and SQL systems for a variety of clients, independently owning reliability with no platform team backstop.

Data Analyst & AI Trainer

May 2023 – Nov 2025

Salt Lake Community College

Applied data science principles — sanitization, ETL, model training — in Jupyter, and engineered SQL/JavaScript-based visualizations.

Earlier Technical Roles

2015 – 2023

Hill Air Force Base · Instructure (Canvas) · Microsoft · Control4

Software development, data, and systems support across enterprise and defense environments.

Full role history on linkedin.com/in/hsmith-dev

Technical Skills

The stack, top to bottom

LLM & RAG

RAG architecture design · prompt engineering · multi-model API integration (Claude, Gemini, OpenAI) · local LLM hosting (Ollama, KoboldCpp) · agentic tool-use orchestration

Modeling & Evaluation

TensorFlow · PyTorch · ML.NET · model training & evaluation metrics · classification models · human-in-the-loop evaluation frameworks

Data Engineering

ETL pipeline design · data sanitization · feature preparation · SQL / BigQuery · Python (pandas, Jupyter)

Engineering & Deployment

Python · JavaScript · C# · REST API design · Git · GitHub Actions CI/CD · Google Cloud Platform · Amazon Web Services · Linux

Availability

Teaching AI, and building it into your product

Teaching & Training

I've spent years as an AI Trainer at Salt Lake Community College and bring that same teaching instinct to workshops, mentorship, and technical training — making RAG, LLM orchestration, and applied ML approachable for engineers and teams who are new to them.

AI Feature Development

I design and build AI capabilities into existing software — RAG search, LLM orchestration, evaluation pipelines, and data infrastructure that makes AI features trustworthy in production, not just in a demo.

Let's talk about your project

Get in Touch

Have a project, class, or team to talk about?

Based in Ogden, Utah — open to AI development work and teaching engagements, remote or local.

Email hello@harrisonsmith.ai LinkedIn linkedin.com/in/hsmith-dev GitHub github.com/hsmith-dev Résumé Download PDF