Work · 2022–present
Oak Grove
Solo operator · US, UK & Japan storefronts · AI-assisted production workflows
Oak Grove is a business I built to answer a specific question: how far can one person, working with AI tools, get toward running an operation across every function a staffed business would need to cover?
Oak Grove is built around 19th-century Japanese katagami stencil designs, adapted into home-decor products. I handle sourcing, design, photography, copywriting, translation, marketing, analytics, and operations, while fulfillment partners handle production and shipping. Over four years, Oak Grove became a production environment for learning where AI can meaningfully support business workflows, where it introduces risk, and where human-in-the-loop review is needed to protect quality, accuracy, and brand trust.
Topics
The brand
Authentic Japanese stencils, built into a home goods brand
Oak Grove transforms authentic 19th-century Japanese stencils into woven jacquard blankets, throw pillows, and other home decor goods. Each pattern begins as katagami: a hand-carved stencil used for dyeing traditional textiles like kimono, furoshiki, and nōren.
After launching in the US, the brand expanded to the UK (2024) and Japan (2025). Everything else was built from scratch: brand identity, full product line, multi-market Shopify infrastructure with English and Japanese storefronts, and a sourcing and curation pipeline for stencil artwork.
2022 – 2025
Phase 1: learning what AI could and could not be trusted to do
From 2022 through 2025, I used AI entirely through chat interfaces: ChatGPT, Claude, and Gemini. Some tasks benefited immediately, including product descriptions, ad copy, translation drafts, market research, and analytics interpretation. Others required caution. AI advice on paid ad setup was often confidently wrong, while primary source documentation was more reliable. Generative image tools were not usable for product marketing until the outputs looked close enough to what customers would actually receive.
That period taught me the first rule of the workflow layer I built later: AI was useful only when I understood the task well enough to judge the output.
↗ Me, Myself, and AI: Building a One-Human, AI-Collaborative Business Operation — LinkedIn article (published March 2026)
2026
Phase 2: turning useful AI tasks into repeatable workflows
In 2026, I moved from ad hoc prompting to structured AI workflows in n8n. Chat had been useful, but it was still manual: one task, one prompt, one output at a time. n8n let me turn repeatable decisions into pipelines, connect AI outputs to Airtable and Shopify, and add human review gates before anything reached customers.
Some workflows saved time. Others made new processes possible, like staging products into rooms from inspiration photos or capturing color palettes from images and applying them to stencil designs. The important design work was not just connecting APIs. It was deciding where AI could act independently, where it needed constraints, and where I had to stay in the loop.
Workflow 01
Inspiration Stager
Before
Writing staging prompts meant starting from scratch each time, describing lighting, furniture, geometry, and mood in text, with no guarantee the output matched what I'd imagined.
After
I upload two images: a room photo I like and a photo of my product. Gemini uses the room photo to understand the lighting, furniture, and layout. Then it places my product into a new image based on that room, while keeping my product's pattern and colors intact.
Inspiration Stager Workflow
Workflow 02
Lifestyle Mockup Generator
Before
Generated lifestyle images one at a time in the Gemini chat UI, manually writing a prompt for each room setting, then repeating. Time-consuming and inconsistent across products.
After
One trigger run produces four room variants from a single sample photo. Room-specific prompts are constructed automatically per product type. All outputs land in Airtable for review before use in listings or ads. The value was not just speed: it gave each product consistent visual coverage across room types while keeping selection and publication under human control.
Mockup Generator Workflow
Input: sample photo I took
Output: mockups for each of 4 scene variationsWorkflow 03
Colorway Pipeline: Three Linked Workflows
Captures color inspiration from product photos, applies it to a stencil to generate up to 6 colorway mockups, then routes the results through two web apps — one for my own first pass, one for external feedback — before anything goes to production.
Before
Chose colors by instinct, like using a color picker as a Ouija board, guided by gut feeling with no record of what I'd tried and no way to compare options.
After
- Save a photo of a color scheme I like, from desktop or my phone.
- Its colors are extracted to hex codes in Airtable.
- Those colors generate pillow mockups in up to 6 colorways.
- The promising ones go to reviewers to vote on before anything reaches production. (not illustrated above)
Step 1 of 3 · Palette Lab
Capture color inspiration from anywhere
Saves color inspiration from photos into a reusable palette record. Photos submitted from desktop or mobile are analyzed by Gemini, which extracts the hex codes and writes them to Airtable, ready to be used by the Colorway Generator.
Palette capture workflow
Grid view
Gallery viewStep 2 of 3 · Colorway Generator
Apply a palette to a stencil and generate up to 6 variants
The Colorway Generator takes a validated palette from Palette Lab and a black-and-white stencil design, and generates up to 6 color mockups so I can make a first pass on which combinations are worth pursuing. Exact color still happens in Photoshop — Gemini gets close enough to evaluate, not close enough to ship.
Colorway generation workflow
Step 3 of 3 · Design Review
Curate a batch, send to reviewers, collect yes / maybe / no votes
Two small web apps I built with Claude Code. In the First Cut web app, I review the finished colorways and pick which ones are worth showing. The second app, Design Review, goes to family reviewers, who vote Yes / Maybe / No on each design and can leave a note. Every vote is saved automatically to Airtable, so I can see what resonated before committing anything to production.
Feedback loop workflow
Curator view: select colorways to send for review
Reviewer view: Yes / Maybe / No per designSupporting workflows
Additional automation across content, publishing, and operations.
What this demonstrates
What four years of solo operation taught me about working with AI
Oak Grove taught me that the hard part of using AI in production is not generating outputs. It is knowing when an output is good enough to trust, when it needs review, and when the task should not be automated at all. The goal from the start was to learn how far AI could stretch one person running a real business, and the answer kept changing as the tools changed.
The risks were specific. AI guidance on paid advertising could be confidently wrong: wrong campaign settings, incorrect setup advice, and wasted spend before I caught the mistake. Generative image quality required constant monitoring because almost-real product images could make customers question whether the thing they ordered would match what they saw. Knowledge was also fractured across separate chat interfaces, with no continuity between them.
Every human review gate in the 2026 workflows exists because of a failure mode I encountered directly. Nothing publishes without passing through Airtable for approval, because I learned exactly what happens when AI output moves too close to production without review.
Related work