An operator who puts AI into production.

I find where a business is losing time to manual work, design the AI system that fixes it, and ship it to run reliably in the background.

These are projects I built as a one-man team with frontier models. They are live and in use. They show the kind of work I can produce.

Looking to join a team doing AI implementation or AI operations. Open to consulting for small businesses. Based in Washington State: remote first, glad to be on site and to travel.

a week of team work taken over by the system
~40 hrs
a week of team work taken over by the system
jobs completed by AI workers, each proven done
2,176
jobs completed by AI workers, each proven done
work items shipped through the controller
529
work items shipped through the controller
of brokers, 2021 to 2025
Top 3%
of brokers, 2021 to 2025

Numbers as of November 18, 2026.

Proof

Five systems, live and in use

01Running daily since September 2026

Omnius: an AI operating system that runs my businesses

One person running a real estate team, a new education company and a video channel. AI chat sessions forget everything and can't be trusted to finish a job unattended.

saved per week
~40 hrs
saved per week
jobs completed
2,176
jobs completed
scheduled jobs
45
scheduled jobs

Supabase Postgres · Claude Code (headless) · launchd · GitHub Actions

Read the full case study

What I built

  • A control plane in Postgres and a crew of AI workers on a Mac mini. Work is filed as items with an acceptance test, and a controller dispatches each one as a signed ticket.
  • Each worker does its job on its own branch. A job counts as done only when a command proves the result exists, the tests pass and a second model has reviewed the change.
  • Scheduled agents handle mail, showings and lead response, and a Chief of Staff report four times a day leaves me one list of what needs me.

How it flows

  1. Spec + acceptance test
  2. Control plane (Postgres)
  3. Controller
  4. AI worker, own branch
  5. Production
Gate: Verify command + CI + second-model review

How it runs

A Mac mini in my office. The controller checks for work every 60 seconds and only wakes a model when there's something to decide. A usage guard stops AI work before it can run into overage billing, and an independent Guard agent reads the other agents' work against their written instructions.

Results

  • Took the place of a personal assistant. The team no longer uses a transaction coordinator, and no contracted web development, marketing or video editing has been needed.
  • About 40 hours a week saved, plus the cost of occasional subcontractors and full-time transaction coordination.
  • 2,176 jobs completed by AI workers since September 10, each closed by a passing check, not by the agent's say-so.
  • 653 work items through the controller: 529 done, 53 failed. Failed jobs are caught and reported, not hidden.
  • 326 second-model code reviews, 55 Guard passes and 299 recorded decisions, so the system knows what was ruled and why.

Still open

The second-model reviewer runs in shadow mode: it reports, but it doesn't block a merge yet.

02Live with real clients

Lead response and showings for Team PNW

Zillow leads go cold in minutes, showing requests arrive by email at all hours, and anything written to a buyer has to comply with fair housing law.

answer to every new lead
Immediate
answer to every new lead
showings handled
101
showings handled
fair housing rules on every message
22
fair housing rules on every message

Claude · rules engine (no model) · Supabase · Gmail + Google Calendar

Read the full case study

What I built

  • A lead bot that reads every email reaching the team inbox, gives each one exactly one class, and writes the first introduction in my voice. The welcome email goes out within moments.
  • Every outgoing message passes a 22-rule fair housing check tied to federal and Washington law. Anything flagged becomes a draft and a tap on my phone instead of a send.
  • A showings pipeline that turns each ShowingTime notice into one calendar event: matched on the appointment number, updated in place on a reschedule, removed on a cancellation, colored by county so each broker sees his own.

How it flows

  1. Team inbox
  2. Classifier
  3. Writer (my voice)
  4. Fair housing check
  5. Send, or draft + tap
Gate: Showings: read the calendar event back before marking done

How it runs

The inbox poller runs every minute. Each lane has its own on/off switch in the database and ships off until I turn it on. An unclear showing notice is never guessed; it's flagged for a person.

Results

  • Every new lead is answered immediately and consistently, day or night.
  • Bot-introduced leads have become dozens of showings and several clients.
  • 101 showings processed since September 23, 92 with a live calendar event; 919 inbound emails classified and logged.
  • A book of about 4,500 contacts sorted into ten categories by AI in one pass, with 122 set aside for human review.

Still open

Zillow status updates run through a browser agent with hard stop rules (no CAPTCHAs, no sign-ins, no account settings), so some steps still wait for a person.

03Launched October 2026

Best Practice: exam prep and a state-approved real estate school

Washington real estate students prepare with dated materials. I wanted a modern product: practice exams, video lessons that fix weak spots, and the 90-hour pre-licensing course. That normally takes a team.

first commit to on sale
7 weeks
first commit to on sale
DOL-approved courses
2
DOL-approved courses
commits
1,351
commits

Next.js · TypeScript · Supabase · Stripe · Mux · Vercel

Read the full case study

What I built

  • A full production web product: a free 50-question diagnostic, a paid question bank and simulated exam, a 75-lesson video library, and an AI tutor that explains a missed question and points to the lesson.
  • Stripe checkout with sales tax, and a school side with seat-time tracking, identity checks, final exams and certificates.
  • Every change built by an AI worker on its own branch; it reaches main only after the test suite, CI and a merge gate pass. Production is released only from main.

How it flows

  1. Work item
  2. AI worker on a branch
  3. CI: types, lint, tests, migrations
  4. Merge gate
  5. Production

How it runs

Vercel with separate staging and production, row-level security on every table, and an hourly monitor that checks the payment webhook and pushes my phone if it fails.

Results

  • Built in seven weeks, September 27 to October 27, 2026: 1,351 commits and 177 database migrations.
  • Washington DOL-approved real estate school (School ID 26027097). Two pre-licensing courses approved November 12, 2026: Real Estate Fundamentals (C26030240, 60 hours) and Real Estate Practices (C26030238, 30 hours). Enrollment opened November 14.
  • The course filing was 2,582 pages across two packages, assembled by the same system.
  • Live and selling at youbestpractice.com and bestpracticewashington.com.

Still open

It's early: paid advertising is paused while the landing pages are redesigned.

04In production

An automated lesson video pipeline

A course needs hundreds of narrated, animated lessons. By hand that's weeks per lesson, and AI narration alone drifts out of step with the screen with nobody catching the glitches.

long-form videos on YouTube
164
long-form videos on YouTube
lesson national course
75
lesson national course
masters in the state courses
90
masters in the state courses

Remotion · ElevenLabs · faster-whisper · ffmpeg · Mux

Read the full case study

What I built

  • A pipeline that takes a script and produces a finished 1080p lesson, narrated in an AI clone of my own voice.
  • Speech recognition timestamps every word, and each on-screen element lands on the word that names it.
  • Automated gates fail a video for a clipped word, a bad splice, low bitrate or wrong loudness. A lesson that fails is rebuilt, never patched.

How it flows

  1. Script
  2. My cloned voice
  3. Word timestamps
  4. Animation on the word
  5. 1080p render
Gate: Audio, picture, encode and loudness gates: fail = rebuild

How it runs

Batches run on the Mac mini, one module per render slot, with a driver that sets a bad item aside and keeps going. A nightly job stages the next day's uploads.

Results

  • 164 long-form lessons and reviews on the @PassYourExam channel, released on a daily schedule.
  • The 75-lesson national course, live in the product, plus the video for two state-approved courses.
  • About 625 video compositions in the project.

Still open

The voice is an AI clone of mine, and I say so. The channel is new; audience growth is the next job.

05Running

Trust layer: a multi-model quality gate and a memory that lasts

AI-written exam questions can be subtly wrong, and a wrong answer key in a licensing course is a real harm. Separately, AI sessions start from zero every time.

course items gated
3,196
course items gated
blind models + a referee
2
blind models + a referee
use of the memory
Daily
use of the memory

OpenRouter (Qwen, MiniMax) · Claude · pgvector · MCP

Read the full case study

What I built

  • A question gate: two different outside models answer each question blind. If both match the key and six rule checks pass, the item is clean; anything else goes to a Claude referee. Any edit sends the item back through.
  • Open Brain: a memory every AI tool I use can search, through a Model Context Protocol server. Search blends meaning, exact words and recency.
  • I talk into a Plaud note pen or dictate; the note becomes a structured record, and proposed actions (a task, a calendar entry, a client note) arrive as approve-or-reject cards on my phone, each with its full trail.

How it flows

  1. Question
  2. Two models, blind
  3. Both agree + rules pass?
  4. Referee if not
  5. Approved bank

How it runs

Open Brain is always available and in use every day. Automatic execution is deliberately off: every action from a voice note is approved by a person.

Results

  • Final exam and module test banks for both state-approved courses gated item by item: 2,386 module-test items and 810 final-exam items, each with a verdict.
  • A referee sweep of 1,208 items found 100 to fix and 2 to fail; 198 flagged items were rewritten and passed a second cold check.
  • The outside-model panel cost $2.10 for 265 items.

Still open

Auto-execute stays off until the system has earned it.

How I work

Plan, orchestrate, verify, ship

  1. Step 1

    Plan

    I write the spec and the acceptance test before anything is built.

  2. Step 2

    Orchestrate

    A controller hands each job to an AI worker on its own branch, with the right model for the job.

  3. Step 3

    Verify

    Nothing is done because an agent says so. A command has to prove the result exists, CI has to pass, and a second model reviews the change.

  4. Step 4

    Ship

    Gated merge, production release, and monitors that tell me when something breaks.

On not having a CS degree

I'm not a software engineer, and I don't write code. I design the system, write the specs, direct AI coding agents, and build the checks that prove the work is right before it ships. Every system on this page was built that way, by me, working with frontier models, almost entirely Claude and Claude Code. What I bring is 20-plus years of knowing how a business actually runs, where the hours go, and how to bring complex systems together.

Stack

What these systems run on

Models and agents
Claude (Opus, Sonnet, Haiku) through Claude Code and headless sessions; OpenRouter for outside models; a Model Context Protocol server; agent skills
Data
Supabase: Postgres, row-level security, pgvector, Vault, edge functions
Product
Next.js, React, TypeScript, Tailwind, Vercel, Stripe, Mux, Resend
Media
Remotion, ElevenLabs voice cloning, faster-whisper, ffmpeg
Operations
GitHub Actions, branch protection, launchd on a Mac mini, Gmail, Google Calendar and Telegram integrations, Zapier, 1Password service accounts

Background

Twenty-plus years running the systems behind a business

  1. 2021 to 2026

    Managing broker and co-lead, Team PNW

    Top 3% of brokers, 2021 to 2025. Sales, clients, compliance, and a team that runs on follow-up.

  2. 2014 to 2019

    Operations owner, Kush Farms

    Ran a regulated grow end to end, including seed-to-sale tracking.

  3. 2000 to 2014

    IT director

    Including the IT behind Willamette Valley Nutritional Products, a regulated supplement manufacturer: production equipment (encapsulators, tablet presses, bottle fillers, heat sealers) and the record-keeping and compliance that come with it.

I've run the systems behind regulated manufacturing and a regulated grow. Drop me into a small business and I know where to look.

Contact

What I'm looking for

I'd most like to join a larger team, where I can learn from people with more experience and work on more complex projects.

I'm also open to consulting: bringing a brokerage online with the tools I built for my own (a CRM and email assistant, transaction coordination, an operating system like Omnius), or similar work for a small retailer or manufacturer, about 50 people or fewer. I'm not looking to run an agency.

Depending on the position, I'd hand off day-to-day leadership of Team PNW and be more selective about the clients I handle myself.

Remote first, and glad to be on site and to travel. Relocation only for the right role, with relocation support, and likely splitting time with home.

hello@shannonframe.com