Builder-Educator · Fuquay-Varina, NC

I build AI systems, then teach people to use them.

As a non-engineer, I direct Claude, GPT, and Codex to ship production software, then test whether it actually works and teach people to use AI with real skill and judgment. The organization I run operates on systems I direct and build this way; each one below is marked live or in build.

2
frontier models cross-check Shield (Claude + GPT)
5
production AI systems built
11,000+
people reached across 19 states
7 in 10
learners adopt a conversational AI tool
What I've built

Systems that are real and running

Each was built by directing Claude and Codex, tested in context, and measured in use where live. Every system below is marked live or in build.

GOLD platform homepage
GOLD · independent 501(c)(3) nonprofit

The GOLD platform

The production web app behind GOLD (Growing Older Living Digitally), a nonprofit that teaches older adults to use everyday technology with confidence. It brings content, workshop registration, outcome measurement, and AI-assisted staff tools into one system; I direct its build and run it.

Next.jsReactTypeScript Sanity CMSVercelAnthropic API
live-gold.org →
Live · try it yourself

S.C.A.M. Shield

Paste a suspicious text or email and it flags whether the message shows scam warning signs, in plain language, with the reasons it flagged and the trusted sources behind them. It runs Claude and GPT as two independent readers over retrieval across a vetted corpus (FTC, FBI IC3, AARP, BBB, and GOLD's own scam framework): both judge the same evidence, and when they disagree the system flags it for human review instead of averaging it away. Built by directing Claude and Codex, with uncertainty flags and an up-front AI disclosure.

323
messages tested across two labeled eval sets
2
independent models cross-check every verdict (Claude + GPT)
100%
scam recall: flagged every known scam in the sets
5
trusted sources + GOLD's scam framework in the corpus

Every message is checked against GOLD's own scam-detection framework and trusted sources (FTC, FBI IC3, AARP, BBB). I evaluated it against the live production route, not a mock, on two labeled sets: a 100-case set (50 scam, 35 legitimate, 15 ambiguous) and a harder 223-case set generated by a separate model. Scam recall means every message that was actually a scam was flagged; the two-model check also caught scams a single model alone would have missed.

ClaudeGPTRAGBedrock Titan embeddingsNext.js
S.C.A.M. Shield result: strong warning signs with reasons S.C.A.M. Shield try-it input
Reliability-focused

Production AI pipelines inside GOLD

Three systems that turn real-world input into structured, reviewed output, each with provenance, human checkpoints, and safe fallbacks.

Paper survey scanner
paper survey → data

Survey scan → structured data

Photograph a stack of paper surveys; Claude reads them into structured records, and a human checks the read-back before anything is filed.

Spoken debrief turned into a structured report
voice note → report

Spoken debrief → structured report

A three-minute voice note from staff becomes a structured report (what they noticed, barriers, follow-ups) matched against what attendees said.

Scam alert rewritten before and after
FTC alert → plain English

Scam alert → accessible rewrite

Federal scam alerts rewritten into clear language for older adults, reviewed by an editor, and published live on the site.

In active development

Karat: AI voice hotline

The toll-free Amazon Connect line is live; the AI phone agent (a Nova Sonic voice model and a Claude-on-Bedrock chat assistant) is in build, running inside a GOLD-owned AWS environment, to extend GOLD's help beyond the workshop room.

Amazon ConnectNova SonicClaude on BedrockAWS
Alara Group website
Alara Group

Alara Group (REPCAL LLC) is a training-program consultancy I'm a member of, separate from GOLD. I built its website in Next.js, engineered for AI discoverability with llms.txt and JSON-LD structured data.

alaragroupus.com →
Also built: Marketing automation: Google Ads + Meta, Python on GitHub Actions Borumi: MCP workflow across Claude, ChatGPT/Codex & HeyGen
Impact & teaching

I measure what people can do afterward

Reach is easy to claim. What I care about is what people can do afterward, so I built the measurement into the platform: bilingual pre/post instruments, task-based observation, and A/B survey variants that separate what people can do from what they say.

85%
complete a real task themselves
90%
feel safer and more prepared for scams
83%
complete a key MyChart task
7 in 10
go on to use a conversational AI tool

From participant surveys and program records · figures as of October 2026

From practitioners to the public

I coach practitioners and staff on prompt chaining, tool orchestration, and output evaluation, and I run a train-the-trainer program across 40+ partner sites. Teaching a skeptical general public is my proving ground for clear explanation, live facilitation, and measurement, and I adapt those disciplines to the more complex workflows advanced users need. I teach it live and measure whether people apply it.

Watch: "What is AI?"
A plain-language introduction for a general audience.
Play the clip →
About

RJ Reyes

Founder and lead educator of GOLD, a Master Trainer, and the person who directed the build of every system here. I started as a ChatGPT user; today AI is the operating system for an organization I run, and my work is helping other people make that same jump: from using AI to look things up to using it to build and decide.

I own the full arc: I design the learning, direct the software that delivers it, and measure whether it works. B.S., Marketing (Coastal Carolina University) and PGA Professional Golf Management.

Email rj.reyes@live-gold.org
LocationFuquay-Varina, NC
See a system in action →