AI safety and safeguards
Testing how models misbehave: adversarial prompts, misuse paths, and safeguard review.
About Jonathan R. Reed
The sociology student who breaks language models, and why those go together.
Dallas, TX | UT Dallas
B.A. Sociology, The University of Texas at Dallas, expected 2027
Open to remote and Dallas-area roles
Jonathan R. Reed
I study Sociology at UT Dallas and red-team language models, a combination that makes more sense than it sounds.
I started poking at LLM behavior in 2019, with GPT-2. Nearly every failure I've found since, from jailbreaks to injection paths to safeguard gaps, traced back to the same place: not the weights, but the users, incentives, and systems wrapped around them. Sociology is the study of exactly that.
Hello.World Consulting is where the testing becomes applied work: red teaming, scoped deployment, private AI workflows, architecture review, documentation, and a handoff the next person can actually run.
At Reed & Terry, a law firm, I handle IT and security: networks, backups, access control, and AI tooling that has to respect confidential client data.
At Podium Education, I teach working AI use to 400+ students and built feedback methods that cut grading turnaround by 31%.
Before any of this I was an EMT. Emergency settings teach you to decide calmly, document everything, and execute under pressure. Those habits transfer directly to incident response.
Public work includes AI Stats, TRACED, Internet Outage Atlas, AI News, and local-first RAG tooling.
Right now I'm looking for internship and entry-level work where models meet real users: red teaming, safeguards, deployment.
Operating habits
-
Tight feedback loops
I show work early. Wrong directions die faster that way.
-
Written handoff
Every project ends with notes the next person can run with, not a folder of mystery files.
-
Private by default
I use local and private-cloud patterns when the risk calls for it.
-
Low meeting overhead
Short written updates instead of standing meetings. Decisions get written down so they hold.
What I Work On
I keep the main site focused on identity, resumes, writing, and selected proof. Deep case studies can live elsewhere when that system is ready.
Trust & safety
How people abuse systems, who gets hurt, and who responds when it happens. The design work, not the theater.
AI deployment
Local-first AI and RAG systems that run on your hardware, documented for the team that inherits them.
Software and security tools
Full-stack tools with security in the design from the start, plus the teaching that gets people to use them.
Why Sociology Helps
Whether an AI or security system works outside a demo gets decided by people: their incentives, their trust, their workarounds. Sociology is the study of exactly that, and it keeps my technical work pointed at the failure modes that show up in production instead of the ones that only exist on benchmarks.
Current Resume Versions
One work history, tailored per opening. The hub keeps a version per track so you can grab the one that matches.
View resume hub