It's Hopper Time What I built in 2026
Marc Hopper

Marc Hopper

I am not an AI guru. I am an operations guy who takes a real workflow, ships a working AI system on it, and then measures it.

COO, Veterans in Media & Entertainment U.S. Marine Corps veteran MBA, USC M.S. AI & Machine Learning, in progress Los Angeles

The work

What I put into production

Everything here is something I designed, built and put in front of real users this year. Each card says what it does, whether it is live, the technique underneath it, and one measured result. Where a project has no number worth quoting, it does not get one.

Veterans in Media & Entertainment

Nonprofit serving 6,600 film and television professionals. I am the COO.
Internal system
no public screenshot

Operations Platform

Live

One system that runs the whole organization: membership, events, check-in, surveys, reporting and the job board. It replaced a stack of disconnected vendor tools.

How it worksNext.js on Vercel, Claude for the generative surfaces, Redis for state
Scale98 pages, 263 API routes, ~2,800 automated tests
The VME job board, live at jobs.vmeconnect.org

Job Board with AI Resume Screening

Live

Members apply to entertainment jobs with no sign-in. An AI reads every resume against the role rubric and flags fit before a human opens it.

How it worksClaude reads the PDF itself and scores it against a written rubric, advisory only
Result30 applications on day one
See it live
Internal system
no public screenshot

Aegis, Continuous Security Testing

Live

Scans our code and live sites on a schedule, then sends an AI attack agent at our own chatbots with prompt-injection payloads to see what leaks. Built because the commercial equivalent was out of budget.

How it worksScanner suite plus a Claude attack agent with canary detection, capped on spend
Result255 findings baselined on the first run, 64 open today
Internal system
no public screenshot

Member Email Platform

Test mode

Replaces a paid vendor for member email. Every send passes a suppression checkpoint seeded with six decoy addresses, and the campaign refuses to run unless all six were caught.

How it worksOwn send pipeline with a warm-up ramp, bounce handling and a proof-of-suppression gate
Scale6,600-member roster
The short film competition site

Short Film Competition Platform

Live

Runs a smartphone short-film competition end to end: filmmaker registration, a protected portal, judge scorecards and a screening event. An AI assistant answers rules questions in the filmmakers' own words.

How it worksMagic-link access with separated permissions, plus an assistant grounded only in the published rules
Result390 of 390 guard tests green at launch
The plain-English event intake form

Events Automation

Live

Paste a plain-English description of an event. An AI pulls out the date, time and format, drafts the copy, then the system creates the event in the membership platform and renders the banner.

How it worksLLM extraction into a strict schema, then a state machine that writes to the CRM
Built onA taxonomy derived from 1,020 real events
The VME member community sign-in gate

Member Community and AI Concierge

Live

A members-only community behind an email-verified gate, with an AI concierge that answers questions and routes real help requests to a person instead of guessing.

How it worksOAuth membership gate, assistant hardened against forged conversation history
Scale18 channels, membership re-synced every 3 hours
See it live
The published volunteer service record

Volunteer Service Record

Live

Reconstructs a defensible count of volunteer hours from eleven separate evidence streams, so the board can report contribution with proof instead of estimates.

How it worksMulti-source extraction and classification, with overlapping time unioned rather than added up
Result611 hours evidenced across 217 active days
See the record

DezLin

Multichannel e-commerce brand. I am co-founder and fractional COO.
The DezLin storefront, where the support agent runs

Customer Service AI Agent

Live

Answers customers on the storefront around the clock. It proves who someone is before it will discuss an order, looks the order up, repairs broken carts, and hands off to a human when it cannot help.

How it worksA Claude agent with real tools, answering from 1,353 approved answers mined out of 9,319 actual customer conversations
How I proved it0 false identity accepts in ~25,000 attack cases; a 40-agent red team found 1 critical issue, fixed and re-tested
Talk to it on dezlin.com
The DezLin wholesale application form

Wholesale Approval Automation

Live

A wholesale buyer applies on the storefront, the application is scored and posted to Slack, and one tap approves or denies it, which fires the right customer email and account tags automatically.

How it worksScored intake, Slack approval cards, and a watchdog that repairs its own failures
Scale1,732 wholesale accounts; retired a paid app
Internal system
no public screenshot

AI Product Photo Studio

Deployed

Upload one product photo, get clean white-background and lifestyle images back, then push them to the right listing. Built so a non-technical family member can run it alone.

How it worksA generative image model does the photography, Claude writes the product copy, the app publishes to the store
Internal system
no public screenshot

Resale Certificate Review

Live

A wholesale customer uploads a tax exemption certificate. An AI reads the document, checks it against that state's rules, and applies or refuses the exemption.

How it worksClaude reads the uploaded document and returns a structured decision that is binding, not advisory

It's Hopper Time

My own channel. Built to prove the pipeline end to end.
A finished vertical short from the video pipeline

Video Production Pipeline

Built

A phone video goes in. A finished, captioned vertical short comes out, with captions written separately for each platform, and nothing posts until I approve it in Slack.

How it worksWord-level transcription, an AI cut list checked against a schema and snapped to word boundaries, one render pass, audio mastered to broadcast loudness
OutputA finished short plus 7 platform captions per video
Internal system
no public screenshot

AMC Movie Release Radar

Live

I pick which upcoming movies I care about and it texts me once, the moment tickets go on sale at my theatres. I applied for and was granted an AMC developer key to build it.

How it worksAMC developer API, a scheduled monitor with a discovery gate, delivered straight to iMessage
Scope5 theatres watched
HopCal, the self-hosted booking page

HopCal

Live

My own booking system instead of paying for one. It checks both my real calendars, books the slot, creates the video link and sends the reminders.

How it worksThe database itself refuses overlapping bookings, so a race condition cannot create one
Cost$0 per month to run
Book a time on it

The offer

What I would build for a studio first

Not a platform. One workflow, one department, eight weeks, with a number at the start and a number at the end.

An AI first pass on an inbound queue

Pick one queue where things arrive faster than people can read them. Submissions, coverage, talent, vendor intake, internal requests. The work is not deciding. The work is reading everything so a person only has to decide.

  1. Weeks 1 and 2 Sit with the team that owns the queue and write down the rubric they already carry in their heads. Measure what time to first response looks like today.
  2. Weeks 3 to 6 Ship a first pass that summarizes, tags and ranks every item against that rubric. A person approves or overrides each one. The AI never has the last word.
  3. Weeks 7 and 8 Measure again. Publish both numbers, including the overrides, because the overrides are what tell you where the rubric was wrong.

Why this one

It is reversible, it touches no customer, and a person sees every output before it counts. If it fails, the queue goes back to the way it worked last month and nobody outside the department notices.

I have built this shape twice

An AI resume screen on a live job board, and a casting submission pipeline for a scripted series. Both shipped. They are the same machine pointed at a different queue.

The honest risk

It fails when the rubric is not really written down and every reviewer is quietly using their own. That is worth finding out in week one, and it is worth knowing either way.