The shop where customer response never clocks off.
gerboni is a live online shop, run by one person, with a customer-response agent built into the storefront: software that answers product and order questions itself, at any hour. This page is the full write-up: what's code, what's AI, what stays human, and what it costs to run. It's the proof behind the customer-response build I offer.
gerboni is my own shop, not a paid client engagement. I built it, I run it, and it takes real payments. It's written up here because it's the working proof behind the customer-response build, and every claim below is open to a source walkthrough in the build review.
An online shop is open all night. Its owner isn't. Customer questions don't keep shop hours: "does this come in medium?" late in the evening, "where's my order?" on a Sunday. Every hour a question waits, the order behind it cools off.
gerboni sells t-shirts carrying the coats of arms (ģerboņi) of major Latvian cities. I run it alone. So the usual answers didn't work: answer everything myself at all hours, hire someone for a one-person shop, or accept the leak. The same three non-options every owner-run business recognises.
A complete commerce build with the support role designed in, not bolted on:
- A customer-response agent in the storefront: product discovery, order questions, customer queries, 24/7. A Pydantic AI agent running on Claude Sonnet 4.
- Product recommendations from browsing history: suggestion-only. The customer decides; the model never orders.
- An admin panel for the deterministic life of the shop: sales, orders, products, users, newsletter. One place, no re-keying.
- Stripe checkout, orders in Postgres. Money and records are code.
The split that makes this safe for a small business: AI where judgement is cheap, code where money moves, a human where it's odd.
Deterministic: code, not the model.
Checkout and payment are Stripe. Orders, products and customers live in Postgres. The admin panel reads and writes those records the same way every time. Money never rides on a model's judgement.
Latent: the model.
The conversation. The agent answers product and order questions at any hour and suggests products from browsing history. Suggestions only: it talks, it doesn't transact.
Human: the owner.
Refunds, disputes, and anything unusual come to me. The agent covers the hours and the repetition; I keep the calls that need an owner's judgement.
- Live and taking real payments through Stripe (source: Stripe dashboard).
- The agent answers around the clock, including the hours outside a 09:00-18:00 support desk (source: always-on storefront).
- Recommendations are suggestion-only: the customer clicks, the model never places an order (source: storefront behaviour, open to walkthrough).
- Sales, orders, products and users in one admin panel (source: screenshots on the results page).
Built on my own time and my own stack. It runs on a small, fixed monthly cost: modest hosting plus pay-as-you-go model usage for the support conversations.
One honest note on the build itself: this shop was also my testbed for the commerce stack. A client build reuses the pattern without the experimentation, which is why it fits a fixed 4-8 week scope.
- Log the numbers from day one. Conversation counts and out-of-hours share should be a weekly export, not an archaeology job. It's why two numbers on this page are still marked for counting, and it's a lesson every build I scope now inherits.
- Simpler hosting for a shop this size. Kubernetes is more platform than a one-person shop needs. For a client I'd pick boring hosting and spend the difference on the workflow itself.
Losing enquiries to slow replies?
This is the proof behind the customer-response build: an agent that answers your customers inside your own stack, with the same strict split of what's code, what's AI and what stays yours. Fixed scope, built in 4-8 weeks, handed over with a runbook. Client builds are LLM-agnostic: OpenAI, Anthropic or locally hosted models. Book a 30-min build review and I'll tell you what it would take in your business, or whether it's not a fit.
Book a 30-min build reviewPublished 10 July 2026