Turning a Small Merchant's Own Data into Decisions

What this is

Merchants who sell on online-shop platforms are not short of data. The platform shows them sales and traffic, and third-party analytics can go deeper still. The gap is not data; it is that most small merchants do not know the data exists, do not know how to read it, and therefore price and advertise on gut feel.

I designed and validated an assistant for those merchants. It reads the accounts a merchant has already connected — marketplaces, social channels, and read-only workspace data — and turns that into a decision the merchant can act on. Any action beyond reading surfaces to the owner before it happens.

How I operated

• Made the honest boundary explicit first. The assistant may read what the merchant has connected. Anything that changes or spends waits for the owner. For a small business, that line is the difference between a tool and a liability, so I designed it in rather than bolting it on.
• Refused to let a claim outrun its evidence. A generated conclusion survives only if it traces back to the merchant’s own captured numbers. A deterministic validation layer sits between the model and the merchant.
• Kept the cost honest. Model routing runs under a hard spend ceiling, with the default path deliberately boring, because token cost is the real recurring cost of a product like this.
• Ran the unglamorous reliability work. Self-hosted across separate providers, isolated egress so platform traffic never touches the main host’s reputation, nightly off-server backups, and a monitor that checks real traffic instead of “is the process running”.
• Shipped the parts a merchant touches. Onboarding, the assistant surface, and the approval step, reviewed end to end before calling it live.

Outcome and value

A live assistant that turns a merchant’s own connected data into decisions, and that cannot act unilaterally on the business. The reliability posture is written down, including the single point of failure I chose to accept rather than pretend away. The value is that a small operator gets the decision-support a larger business would build in-house — without handing over control of the shop.

What this shows about how I operate

The hard part of an AI product is not the model; it is deciding what the model is allowed to do and how you will know when it is wrong. I start from the trust boundary and the evidence rule, then build the product around them — the same way I gated releases and reconciled evidence before each release.