rg. Let’s talk

Forward deployed engineering & AI leadership. Dallas–Fort Worth, TX.

Hello.

I’m Reed Garrett. I put trustworthy agents to work inside real operations.

Explore my work
Reed Garrett in a patterned polo
48 hours → 5 minutes

Two days of waiting, gone.One iPad station now does what three teams did.

1 → 5 agents

One win became a program.The first agent sold the next four.

$60M → $288M

AI value, multiplied 4.8×.More value from every model, not just more models.

Selected work

Not demos. Three systems doing real work today.

Three bottlenecks, each one found in person and fixed in production.

  1. 01Work waits in line between teams.
  2. 02The answer is buried in a document somewhere.
  3. 03Pricing lives in one expert’s head.

01Work waits in line between teams.

Two days of paperwork. Now five minutes.

Intake agentIn production
Photograph. Read. Match. Write. Print. One station turns a part and its paperwork into an ERP work order: 48 hours → 5 minutes, 7,341 intakes.

A part used to get handled at least four times before a technician saw it: unboxed, shelved, keyed in, photographed. Up to 48 hours. Now it’s touched once. An iPad at the dock does the photos, reads the paperwork, builds the work order and prints it. Zero lost in 7,341.

How I rebuilt the intake

02The answer is buried in a document somewhere.

26,326 manuals in. One cited page out.

Manual searchIn production
Every answer comes with its page: AI search narrows 26,326 manuals to the one page that answers the job, and cites it.

When the wrong answer can put the wrong part on an aircraft, “close enough” is a failure. The search finds the manual that governs the job and cites the exact page. Its fact-checker ruled on 4,266 part claims across 3 million pages: zero audit failures.

How I built the search

03Pricing lives in one expert’s head.

Process first. AI second. Quotes on autopilot.

Pricing engineIn production
Same part, same job, same price, in seconds: a new job is matched to its last quote, and quotes on target rise from 45% to 91%.

Quoting ran on handwritten notes passed between technicians and quoters, and every question meant another lap. Before any AI, the handoff went digital in a web app built on the ERP. Then 157,061 past quotes set the rules, and any job the data can price with confidence now quotes itself.

How I rebuilt the pricing

Leadership · Fortune 20 telecom

I build teams
that scale AI.

At a Fortune 20 telecom, I built the team that took an AI portfolio from $60M to $288M a year: engineers who embed with the business, ship to production and keep every model earning. Plus a $14M vendor contract cut in half and a patent-pending edge AI method.

How I built the team
Portfolio viewIn production
Annual AI value$288M, up from $60Mup from $60M · 4.8×
163 models · every one monitored and valued
Vendor cost$14M → $7M
Customers saved in a year47,600
Edge AI methodPatent pending

How I work

The playbook behind every number above.

01

Embed with the people doing the work.

Discovery happens on the receiving floor, not in a conference room. The people doing the job show me what no spec captures.

02

Trust the data over the brief.

Asked for faster quotes, the first fix was the process, not a model. Then 157,061 past quotes decided where automation could be trusted.

03

Own the outcome, not the handoff.

The job isn’t done at launch. 99.8% uptime, responses cut from 20 seconds to 9, AI costs down a third, and still improving.

Off the clock

One choir, 420 venues, one night.

I built Watch Party Central end to end. An agent turned host sign-ups into ticketed events several times a day, and fans found a party by ZIP code: 57,190 tickets, 17,431 households, 44 states.

Visit Watch Party Central

Aerospace repair · Intake agent · In production

Two days of paperwork. Now five minutes.

48 hours → 5 minutes per work order

The problem
Nobody was slow. The workflow was. Each part went from a receiving crew that opened, inspected and matched it to its paperwork, onto a shelf, to a data-entry team that keyed it into the ERP, then to a photo station that shot it from six angles and bagged it. Three teams, two queues, up to 48 hours before a technician could touch it.
What I built
I started on the receiving floor, with leadership’s priorities in hand, and rebuilt the whole path around one station. An iPad photographs the part, a vision agent reads the paperwork and data plate, the ERP resolves customer and part, and the finished work order prints right there. I built the app and the AI, trained the crew, and wired it into the ERP with their integration team.
The clever part
One AI call inside a tightly scripted workflow, so it’s fast, cheap and predictable. Anything uncertain lights up red before it can reach the ERP. Launch-week printing got solved by routing through the ERP’s own print command, and smart caching cut daily inference spend by a third.
What changed
A complete intake in five minutes on day one. Within weeks it was the front door for 90% of the shop’s work: 7,341 intakes, zero lost. It runs at 99.8% uptime, with monitoring agents checking its health around the clock. The CIO: “These are unbelievable stats.” One agent became five.

Python · FastAPI · Claude vision · Next.js · ERP integration · Evaluation · On-site monitoring

Back to the work Email me

Aviation maintenance · Manual search · In production

26,326 manuals in. One cited page out.

26,326 manuals → 1 cited page per question

The problem
Every business keeps its answers in documents nobody can search mid-job. Here it was 26,326 maintenance manuals, where a confident answer from the wrong manual can put the wrong part on an aircraft.
What I built
Search that matches each job to the manual that governs it, reads the tables and diagrams, answers in plain language and cites the page. When the manual is silent, it says so instead of guessing.
The clever part
A fact-checker that read 18,939 manuals, pulled 323,000 citations and 84,000 part pairs, and ruled on every one of 4,266 alternate-part claims: confirmed, contradicted or undetermined, with the page and the exact wording.
What changed
4,266 claims checked across 3 million pages at 99% citation accuracy, and a 2,000-row audit with zero defects. Technicians get the answer and the proof in seconds.

Python · Agentic retrieval · Document ingestion · Citations · Evaluation

Back to the work Email me

Industrial services · Quote automation · Live since August 2026

Process first. AI second. Quotes on autopilot.

45% → 91% of quotes on target

The problem
The ask was simple: quote faster. But quoting ran on paper. Technicians handwrote diagnostics, parts and labor; quoters retyped them, chased answers and dug through the ERP for prices. Similar jobs got different numbers, and every question added another lap.
What I built
Process first. A web application on top of the ERP gave technicians and quoters one digital handoff: no retyping, no chasing. Then the pricing logic: 157,061 past quotes backtested to find where history predicts the price, down to a 0.38% median error.
The clever part
Automation only where it’s earned. When a job clears the confidence bar, the quote goes out automatically. When it doesn’t, it lands in front of a quoter with the history already attached. Eight fail-closed gates decide which is which.
What changed
Three quotes went out the first night, and customers responded within hours. 91% of quotes now land on target, up from 45%. In the customer’s words: “If we already have the labor, that’s half their job.”

Web app on the ERP · Python · T-SQL · Backtesting · Confidence scoring · Automated quoting

Back to the work Email me

Fortune 20 telecom · Director, AI/ML & GenAI · 2024–2026

The team behind a $288M AI portfolio.

$60M → $288M annual AI value

The problem
Big companies don’t lack AI models. They lack a team that knows how to turn them into value at scale. This portfolio had 163 models in production and a mandate to grow.
What I built
A team built to scale AI: engineers who sit with the business, ship to production and measure every model in dollars. I hired, coached and led it, built its model operations from zero, and gave the board one view of every model’s health.
The clever part
Every model had to earn its keep, so the team’s time went where the value was. That discipline made the portfolio easy to grow and impossible to argue with. Along the way I co‑invented a patent-pending method that runs AI on idle network compute.
What changed
$60M → $288M in annual value. A $14M vendor contract cut to $7M with every revenue feature kept. 47,600 customers saved in a year. Before that: an experimentation program that delivered $165M against a $50M target, and the company’s first GenAI marketing team, whose AI copy beat human copy by 21.5% and cut go-to-market from 14 days to 5.

AI strategy · Model operations · Experimentation · GenAI products · Governance

Back to the work Email me

What’s your messiest workflow?

Let’s talk.

Tell me what you’re trying to change.

Dallas–Fort Worth · MBA · Patent pending · English & Spanish