# Vu Tran — AI Systems Designer

> Orange County, CA. I design the system, direct AI to build it, and check it before I call it done. The limit on what ships is how well the problem is understood, not how fast anyone types. This page is a running record. It grows every time I ship something or write down a new way of working.

Contact: vu@vtdigitaloperator.com · 949.212.6489 · https://linkedin.com/in/vutran-digital-creative
Updated: 2026-10-06

## Evidence levels
- Spec: Designed and written down. Not built yet.
- Built: Works, demonstrated end to end.
- Running: Live and in regular use.

## Agent systems and automation
Scheduled agents and tool chains that run without me starting them, with a checkpoint before anything leaves the system.
- OC Referral Radar [Running, 2026-10]: Twice a day it reads public local community groups and 11 Orange County subreddits for "who do you recommend?" posts, matches each to a member of my referral group, checks the thread so nobody gets referred twice, and drafts the reply for me to post. Proof: 885 posts read in the first four days; 8 matches, 5 replies posted, 1 duplicate caught; $5.92 in scraping cost against a $19 cap.
- Agent toolbelt [Running, 2026-09]: The services my agents can call directly over MCP and API, so generation, publishing, research and email run as steps in a job instead of tabs in a browser. Proof: HeyGen API wired in, about 90 tools; Magnific, DataForSEO, Apify, Buffer, Kit, Make, Notion, Gmail and Calendar connected; Custom MCP setups for tools with no marketplace listing.
- Hermes agent stack [Running, 2026-08]: An agent stack on OpenRouter that runs recurring research, content and reporting jobs on a schedule. Every job has bounded steps and a human checkpoint before anything leaves the system. Proof: Full stack live in about two hours, Aug 21; Daily 7am job deployed and verified; Skill library under git version control.
Rule: Reported success is not verified success. Every write gets a unique filename and a checksum match before it counts as done. (Origin: A file transfer reported success twice while writing stale bytes. One of those shipped a 53-cue transcript in place of the 145-cue one.)

## Voice and conversational AI
Phone-callable agents that capture a lead mid-call and route it to the owner with nobody touching anything.
- Phone-callable AI receptionist [Running, 2026-09]: Call a real Orange County number and an agent in my cloned voice runs the intake: problem, city, urgency, callback window. Mid-call it fires a tool that posts to a Make scenario, which writes a row to the owner's log and pings the owner's phone. Proof: Live call tested end to end, Sept 5; ElevenLabs + Twilio + Make + Google Sheets + Telegram; 9-field lead log, filled with nobody touching it.
Rule: To guarantee a field gets collected, put it in the tool-call trigger, not in the list of things to ask. (Origin: On the first live call the agent skipped the caller's city to keep the call short. Optional fields get dropped.)

## Web apps and internal tools
Working apps with real data behind them, from a 22,000-song music library to review consoles that hand their output to an agent.
- VTDO Jams, a private cloud music library [Running, 2026-10]: Brought a 10-year-old iTunes library back to life as a private web player with search, BPM filtering, playlists, lock-screen controls, duplicate checks and daily backups. The migration ran read-only with stop-and-approve gates. Proof: 22,649 songs loaded from a 148.6 GB audit; 82 automated tests and 29 browser checks passing; Independent code review found 15 issues, all fixed.
- Cut Notes [Running, 2026-09]: A video review console. I watch a proxy and press M wherever something needs to change; it captures a frame-accurate timecode, tags and a note, and the editing agent works the cut from that list. Proof: 28 notes across two parts of the first episode; Keyboard-driven, single-frame stepping; A "part done" signal so the agent never reads a half-written pass.
- VTDO Work Release [Built, 2026-10]: A thirty-second permission form a client signs on their phone so I can show the work we did. They opt out of anything they'd rather not include, sign with a finger, and both of us get a timestamped PDF. Proof: Four opt-out permissions, from logo to quoted words; Finger signature with date, time and time zone; PDF generated in the browser, copy sent to the client.
- OC PRO member directory [Built, 2026-10]: A phone-first directory for my referral group. Search by name or service, filter by service type or board, then call, text, email, save the contact with a headshot, or send the member's card to someone who needs a referral. Proof: 33 members, built from their public bios; One-tap contact save with photo; Link previews built for texting a card.
- Two production sites in 48 hours [Built, 2026-08]: Research, copy, design, build and launch on two different stacks inside two days. Proof: WordPress with Elementor; Vercel-hosted build.
- Demo Foundry [Built]: A platform for per-prospect demo builds with expiring share links, a Draft to QA Ready to Active to Expired status workflow, and view tracking so I know a prospect opened it before I follow up. Proof: Status workflow designed before any page was built.
Rule: Before saying "fixed," check what the change broke. Forward and reverse, a range of sizes, and the deployed build, not the local file. (Origin: A text replacement matched nothing and reported success. A correct CSS rule got overridden further down the file. A layout checked at one screen height broke at every other.)

## iOS apps
Native iPhone apps built by directing AI, with my own Apple Developer account in place to ship them.
- VT Field [Running]: An iPhone app that turns a business card into a saved contact. Scan the card, say the details out loud, or type them, and it builds the contact for you. Proof: 30+ business cards processed so far; Three ways in: scan, voice and type.
- Workout tracker [Built, 2026-10]: A fitness tracking app for iPhone, built over a weekend. Proof: Built Oct 3 to 4.
- OR Timer [Built, 2026-09]: My first fully functioning iPhone app. Proof: Apple Developer enrollment complete.
- FieldNote [Spec, 2026-09]: An offline iPhone app for field technicians: talk through the job and Apple's on-device model returns a typed job report, priced from the client's own rate table. Includes Swift samples, a phased build plan and an airplane-mode demo script. Proof: Full build spec written.
Rule: On-device AI writes the report. It never sets a price. Prices come from the client's own rate table. (Origin: A design rule from the FieldNote spec: a model that invents a number on a customer invoice is a liability, not a feature.)

## AI video and presenters
A synthetic presenter pipeline driven through the HeyGen API, with measured settings and quality gates that run before I see anything.
- AI presenter pipeline [Running, 2026-09]: Long-form episodes presented by a synthetic version of me, rendered through the HeyGen API with my cloned ElevenLabs voice. The render settings were chosen by controlled A/B tests, not by eye. Proof: First episode rendered: 8 min 21 s, split on beat boundaries; Measured: 3.2x realtime render, about 16 credits a minute; Breath noise gated at -39 dB, sibilance fixed with a de-esser instead of a denoiser.
- Video QC gates [Running, 2026-09]: Two scripted gates check every graphic before render and every rendered file after, then two reviewer agents judge motion and edit. Failures get fixed and re-run before anything reaches me. Proof: Layout gate: clipping, overflow, safe area, palette; Render gate: empty, static, hold, frame rate, duration, duplicates; 33 of 33 cuts snapped to real transcript cues.
- AI-generated video slideshow [Built, 2026-10]: A video slideshow with every element generated by AI. Proof: Built Oct 3 to 4.
Rule: Nothing is timed against an estimate. Every cut snaps to a real transcript cue within 0.30 seconds. (Origin: Graphics timed to a modeled transcript landed 0.6 to 3.0 seconds off, and a render got rejected for it.)

## Image generation and content
Campaign-scale image work run from a reusable spec, plus content pipelines that publish through Buffer.
- Generative image practice [Running, 2026-09]: Campaign work runs off a reusable prompt spec that locks lens, light, palette and a reserved text zone, and varies only the subject. The image platform is wired into the agent stack, so a batch runs with identical art direction every time. Proof: 1,395 assets produced; 1,051 generations, 262 edits, 57 video pieces; Driven programmatically over MCP.
- Social content pipelines [Running, 2026-08]: Content for two brands drafted, rendered and queued to Pinterest through Buffer, with a scripted pin builder doing the layout. Proof: Python pin builder with a Node renderer; Posts staged through the Buffer API.
- Faceless Avatar Operator [Built, 2026-09]: A prompt console that holds a locked identity block for each synthetic presenter and writes it in the dialect each of 20 image and video models expects. Claude vision reads the reference shots and repairs prompts that miss. Proof: 20 models across Photoshop, Magnific, Higgsfield, Gemini, ChatGPT and HeyGen; 3 recurring presenters; Identity derived from 5 to 8 reference shots.
Rule: Lock art direction in a reusable spec and vary only the subject. For a recurring face, keep only the features that hold across every reference shot. (Origin: One-off prompts drift by image forty. An averaged feature is one no reference photo supports, so the model renders it differently every time.)

## SEO and market research
Prospect diagnostics built on DataForSEO and public sources, with free screens that run before a cent is spent.
- Leak Finder [Running, 2026-08]: Finds a specific, fixable, dollar-valued problem in a business from public evidence: the ad library, reviews, missed inquiries and site health. Three free screens run before any paid data pull. Proof: Measured funnel: 30 screened, 12 qualified, 10 live, about 5 worth a paid pull; Three-source gate on every local prospect; Every platform-computed number recomputed from raw records.
- Local ranking grid audit [Built, 2026-10]: A Maps ranking grid for a local contractor: 9 cities by 3 searches, plus a broken main domain, a missing business category and stale reviews, all found from public data. Proof: 27 ranking checks, 21 outside the top 20.
Rule: One source is not a check. When sources disagree, that disagreement is the finding. (Origin: I once cleared 18 leads as "no website" from a single API field. Six of them had live sites.)

## Funnels and marketing automation
Opt-in funnels, nurture sequences and lead flows that run on their own once a visitor raises a hand.
- AI presenter funnel [Running, 2026-09]: A lead-magnet opt-in on my own subdomain with double opt-in, download on confirm, and a five-email nurture sequence behind it. Proof: Live Sept 9; 14-page guide as the magnet; Tag-triggered sequence in Kit.
- Missed-call SMS flow [Built]: When a call goes unanswered, the system texts back, qualifies the caller, writes the answers into fields, assigns a named owner and escalates against a callback deadline. Proof: Registered business texting line, brand and campaign approved.
Rule: Confirm there's a person and an address to reach before writing a word of outreach. (Origin: I wrote five outreach emails, then found there were no email addresses to send them to.)

## Skills, specs and AI enablement
Every working method gets written down as a skill someone else can run, versioned in git, with the numbers that justify it.
- The Blackbook [Running, 2026-07]: My skill library. Every process I work out gets written down with its trigger, its steps, its checks and the numbers behind it, then versioned, so the next run is identical and someone else can run it without asking me. Proof: 43 skills written since July 29; Versioned in git with an index and an archive; Four shared publicly below.
- Conversational lead schema [Spec, 2026-08]: A field schema, write rules and tag system for an Instagram DM lead-capture system, written for a prospect from their public job post. It separates AI-extracted values from confirmed ones so the team keeps trusting the data. Proof: 27 fields, a write rule for each; 4 conflicts found in the client's own field list.
- Mixed-cart order routing [Spec, 2026-08]: Line-item routing for a fulfillment case the obvious design breaks on, plus the system-of-record boundary that keeps the integration maintainable. Proof: Written on spec, before any engagement.
Rule: Taste gets changed. Craft becomes a rule and a check, so it can't happen twice. (Origin: A line-height that clipped letter descenders shipped three times. Now a layout gate fails the build before it can ship again.)

## Shared skills
- Revision Self-Check: https://proof.vtdigitaloperator.com/skills/revision-self-check.md — The check that runs after every code change: did the edit apply, what else could it break, forward and reverse, a range of sizes, the deployed build.
- Leak Finder: https://proof.vtdigitaloperator.com/skills/leak-finder.md — Find a specific, fixable, dollar-valued problem in a business from public evidence, with three free screens before any paid pull.
- Video Pipeline Ops: https://proof.vtdigitaloperator.com/skills/video-pipeline-ops.md — The operating contract for a multi-asset video build: ground truth before authoring, unattended passes, QC gates and the hard stops.
- Blackbook: https://proof.vtdigitaloperator.com/skills/blackbook.md — How I turn any process or hard-won lesson into a permanent, versioned skill.

## Before AI
- Until 2026: Senior Web Designer. A long-term in-house role.
- 2020 to 2022: Team lead and front-end builder, enterprise event work. Led a small team of developers and designers delivering for Live Marketing's clients Boston Scientific, TE Connectivity, BD, Vizient and IDEXX. Designer and primary front-end builder on two event microsites, and shipped a React trade-show app with the team.
- 2015 to 2022: Founder, Climb SEO. Ran my own SEO practice for seven years.
- Before 2015: Lead designer and art director, creator economy. Designed both sides of a creator-monetization platform where thousands of independent broadcasters earned a living: viewer discovery and subscriptions, and creator earnings and broadcast tools. One of the most-visited sites on the internet during my tenure.
- Late 1990s: Graphic designer, Art-O-Graph Publications. Where it started. BA in Visual Communication.

## Sites I've built
- VT Digital Operator (My site): https://vtdigitaloperator.com — My own site. It opens on what a business is losing, then walks through the teardown, the checklist and how I work.
- Specialists Auto Glass (Spec build): https://specialists-auto-glass.vercel.app — Mobile auto glass in Santa Ana. Leads with the mobile service and the shop's Google and Yelp ratings.
- Meridian (Spec build): https://miami-site.vercel.app — A private relocation brokerage for people moving to Miami from New York and California.
- 27 Ocean Crest (Spec build): https://newport-cliffside.vercel.app — A single-property site for a Newport Coast estate listing, told room by room.
- Ritz Nails (Spec build): https://ritz-nails.vercel.app — A two-location Orange County nail salon, with each location's hours, rating and phone up front and the full menu priced.
- AAA Transmission (Spec build): https://aaa-transmission-santa-ana.vercel.app — A Santa Ana transmission and repair shop, built around its perfect 5.0 across 215 Google reviews.
- Panic Room Builders (Spec build): https://panic-room-builders.vercel.app — Safe rooms for estates. A quiet, high-end tone that leads to a private consultation.
- Puppers Paradise Resort (Spec build): https://pprdb-site.vercel.app — Dog boarding and daycare in Riverside, structured around the services and the city people search for.