NoBanks Nearby: Art by a Human, Business by AI

NoBanks Nearby: art by a human, business by AI. WHAT THIS IS I am NoBanks Nearby, an artist working out of a house in the high desert of California. I make the art. AI agents run the business. Not a demo and not a pitch deck: a real operation with real inventory, real strangers as buyers, and platform verified payouts, running on hardware I own. The thesis is simple and it comes from my own life. Solo artists do not burn out on art. They burn out on operations. Listings, copy, pricing, categories, shipping configuration, customer questions, follow ups. I have twenty years behind a camera and a catalog of over a hundred thousand pieces. The operations were the wall I could not climb alone, and I would not have climbed it manually. It would have burned me out. So I built the business as an agent system instead, inside this hackathon window. One thing up front, before any number below. Five completed sales prove the pipeline works end to end on strangers' money. The first four are resale and household inventory that was on hand to photograph while the print catalog was being built. The fifth, on August 13 inside the entry window, is my own original art: a one of one framed print from my Chromatic Spaces series, listed by the engine on August 12 and bought by a stranger about a day later at my exact preset offer floor, prepaid label printed, payout pending platform settlement. The same unmodified pipeline runs the rest of my framed prints, which are listed and live now. You should not have to dig to find any of that. WHAT THE AGENTS ACTUALLY DO IN PRODUCTION The listing engine. Point it at photos of an item. Gemini reads the photographs and describes what is actually in them. A local model then writes the listing: title, description, condition, attributes. The engine proposes the category and snaps it to the marketplace's fixed list, configures shipping including the package weight tier, attaches the media, and publishes the listing end to end. Every run writes a timestamped result line, and every Gemini call writes a receipt to disk. The catalog. 2,727 original pieces, every caption written by AI vision that actually looked at the piece, not by a template. That catalog is the inventory the engine lists from. The naming system. Every piece names itself. Most are named by the first six hex characters of the SHA-256 fingerprint of their own image file, so the name is derived from the artwork itself and cannot collide or be faked. One collection is named differently on purpose: it consists of 1,440 pieces, one for every minute of a day, and each piece is titled by the minute it represents. The operations tail. Shipping labels are prepaid by the buyer and configured by the engine. Packages go out through carrier pickup at the gate, so there are no car trips and no fees I did not choose. The margin stays home. HOW AI IS TRANSFORMING THE WORKFLOW Before: I photograph an item, then spend twenty to forty minutes writing a listing, guessing a category, fighting a shipping configuration, and second guessing the copy. Multiply that by a catalog and the business never happens. That is not a hypothetical. That is why it had not happened for twenty years. After: I photograph an item and hand the folder to the engine. On a clean run the listing is live in about five minutes, written by AI that looked at the actual photographs, categorized, shipping configured, published. Not every run is clean. The marketplace changes its form, and the code carries explicit guards that abort rather than publish something wrong, plus a continuation step for when its shipping picker stalls. Fifteen straight publish attempts failed before I read the captured payload myself and found the cause. I am leaving that in the repo because a system that admits where it breaks is worth more than one that claims it never does. I review and set price. That is my whole job now. I want to be exact about which decisions are the model's and which are not, because a judge can read the source and check. The model decides the content. What is in the photograph, what the item is, what to say about it, the title, the description, the tags, and a proposed category and condition. That is real generation from the actual image, not a template. Deterministic code decides the guardrails, on purpose. Price, the shipping weight tier, the cash-only pickup terms, whether the run publishes or stops, and the final snap of the model's proposed category into the marketplace's fixed list. A comment in the source says plainly that the model is not trusted with the cash-only rule. That was a deliberate trust boundary, not an oversight, and I would rather show it than claim autonomy I did not build. So AI is not assisting at the edges. AI executes the part of this business that used to consume my life, which is the writing and the judgment about the work itself, and every one of those decisions leaves a receipt I can audit. GOOGLE CLOUD AND GEMINI USE The Gemini API is the production vision layer of the listing engine. It reads the photographs of every physical item the engine lists, and a local model writes the listing from what Gemini saw. Local models handle the reasoning and writing so the operation stays affordable and stays mine. Gemini is the part that sees. This is not a demo path. Every listing cooked through the pipeline writes a timestamped receipt recording the Gemini call that read its photos, and those receipts are in the evidence pack. The hackathon manager confirmed in writing on August 11, 2026 that using the Gemini API satisfies the Google Cloud product requirement. REVENUE, STATED PRECISELY Total revenue from arms-length third-party customers during the hackathon period: $259.13 net of platform fees, across the first four completed transactions with four distinct buyers, all strangers. The fifth sale, my original framed print at $100 on August 13, is excluded from that net figure only because its platform fee settles at payout; it is itemized below. Related-party revenue: zero. I want to be exact about what has sold, because precision is the whole brand here. The revenue to date came from inventory the agents listed, priced, shipped, and sold: resale and household inventory that was physically on hand to photograph while the print catalog was being built. Then, on August 13, inside the entry window, the first piece of my own original artwork sold: a one-of-one framed print, listed by the engine the day before, bought by a stranger at my preset offer floor. One original sale so far, stated exactly as it is. I would rather show a judge a small true number than a large soft one. The monthly shape tells the real story. May, June and July: zero. August: all of it. Revenue started the week the agents could publish end to end without a human finishing the form. Marketing and customer acquisition spend: zero dollars. No ads, no boosted listings, no sponsored placement. The AI written listing copy is the acquisition channel. Buyers arrive through organic marketplace search, at a rate of 2,883 listing clicks in the trailing 30 days. TWO THINGS THIS IS NOT This is not dropshipping. Print on demand of my own original artwork is manufacturing my own product on demand. I own the source work. Nothing here is a resold third-party product sourced from a catalog. This is not a marketplace-specific script. The pipeline is channel agnostic by design. It reads photographs, writes copy, proposes categories, configures shipping and publishes. The marketplace it publishes to is a target, not the product; the same engine can drive any channel that accepts a listing. PRE-EXISTING WORK DISCLOSURE Everything specific to this project was created after May 19, 2026: the listing engine, the the AI copy pipeline, the naming and registry system, the caption corpus, and the evidence tooling. The project runs on standard, publicly available foundations that I did not write and do not claim: Python for the engine, the MLX framework for local model inference on Apple Silicon, the Hermes agent framework, which I use as an end user and did not author or contribute to, open source model weights, and the Gemini API. No pre-existing product, template or boilerplate of mine was carried into this project. The business logic, the agent orchestration, the pipeline and the artwork are original to this entry and to me. CATEGORY: ENTREPRENEURSHIP AND JOB CREATION The category asks who gets to be a founder. My answer is that operations are the gate, and AI is the key. I am disabled, neurodivergent, and I have watched creative businesses die on paperwork rather than on talent. The wall is never the making. The wall is the hundred small operational decisions per item that a solo operator cannot sustain and cannot afford to hire out. Remove that wall and a category of people who were structurally locked out of entrepreneurship can operate a real business alone. That is not aspirational here. It is the receipt. A one-person art business, built and operated inside ninety days by agents on a machine in a spare room, that listed inventory, acquired strangers as customers, shipped product nationwide, and was paid. Small numbers, true numbers, and a system built for one seller so far. I will not claim it is multi-tenant, because it is not. What is proven is that the hard part, the operations that kill solo creative businesses, can be run by agents on cheap owned hardware by one disabled person alone. That is the part that generalizes, and it is the part nobody had shown me was possible. I am the artist. The business runs itself.