travis sawyer

AI Enablement · Adoption · Responsible AI

I help organizations put AI to work — and I can tell whether it’s safe to trust.

Two production AI products, shipped solo. A comparative 8-model safety audit. Fifteen years translating technical work for people who don’t speak tech. I close the gap where most AI rollouts stall: getting non-technical teams to actually use AI, and trust it while they do.

Open to remote roles in AI enablement, adoption & responsible AI

Flagship · Model Safety Audit

I continually put frontier and budget models through adversarial safety audits.

“Which model can we actually trust for this?” is the question that decides whether an AI rollout helps or quietly blows up. Most teams answer it with a vibe. I answer it with a method: push each model the way a real user under stress would, and watch for the moment it stops being safe or honest.

I score production models against a repeatable failure taxonomy, making real ship / no-ship calls — rejecting the vast majority before they ever reach a user. The taxonomy is the part that travels to any organization’s use case:

  • Charm trap

    Sycophancy under pressure tells the user what they want to hear instead of what’s true or safe.

  • Half-rescue

    Boundary abandonment starts a safety intervention, then drops it when the user pushes back.

  • Drift

    Contract erosion quietly loses its instructions across a long conversation.

  • Cliff

    Capability collapse misreads intent or nuance and misfires on the instruction that mattered.

13models evaluated
10rejected on safety or privacy
2shipped to production

That’s the discipline I’d bring to your rollout: not “is the AI impressive,” but “will it hold up in front of the people who depend on it” — with a method you can hand to the next person.

Built & Shipped · Solo

Two production products. The judgment is in the choices, not the code.

GnomeOwnerLive

Problem
Volunteer HOA boards — non-technical, personally accountable for other people’s money — need AI help but can’t survive a confident wrong answer.
Decision
AI drafts, humans decide. The AI runs the essential backend legwork so volunteer boards don’t have a second job, but it never takes an autonomous action. Every output is a reviewable draft. Answers cite their source so a skeptic can verify. Data isolated per organization.
Outcome
A live product non-technical volunteers actually trust — because they can see the sources, approve the work, and turn it off. Payments run on Stripe Connect into each board’s own account.

ReGildLive

Problem
Build an AI that stays consistent and trustworthy across different model providers, and prove it before shipping.
Decision
Run it across Claude, Gemini, and GPT, choosing the right model per task and gating each one through the safety audit. Encrypt user data so the user holds the keys.
Outcome
Production platform with response time cut from 14.3s to 3.9s, and all 46 findings from a security review closed, including per-user isolation and key rotation.

Writing · How I Think

Forensics

How I Got Here

Fifteen years reading a room before I ever read a model.

Before AI I was a cinematographer and director of photography — History Channel, Home Depot, Coca-Cola, Ford. The real job was always the same: take a client’s half-formed idea and turn it into a finished thing, on deadline and on budget, while keeping a room of non-technical people aligned and confident.

That is the whole job of AI adoption, just pointed at models instead of cameras: translate, build trust, and get people to actually use the thing. Recent work in 2026 includes color for The King Center and cinematography for a project with the King family and Senator Raphael Warnock. I still do a little of it. Most of my attention is on helping organizations use AI without the parts that blow up.

Get in Touch

Let’s talk about putting AI to work — safely.

I’m looking for a remote role in AI enablement, adoption, or responsible AI: helping an organization choose the right AI, deploy it responsibly, and get its people to trust and use it.