AI features that ship, not AI features that demo.
The gap between an impressive demo and a feature merchants trust is made of unglamorous work: evaluations that catch regressions, a cost ceiling so a viral week does not produce a surprise invoice, a human-review path for anything that writes to the store, and a fallback for when the model is wrong — because it will be.
We have shipped AI into five apps in our own catalogue: tagging rules, SEO fixes, generated social posts and virtual try-on on product pages. That is the experience we bring — not a prototype, but the second month, when the outputs have to keep being right.
What the engagement actually contains.
AI that ships, not AI that demos: tagging rules, SEO fixes, generated social posts, virtual try-on, assisted support replies.
- 01
LLM features with evals and cost ceilings
A test set that runs in CI, so a prompt or model change that degrades quality fails before merchants see it — plus per-tenant spend caps and visible usage.
- 02
Retrieval over your catalogue and history
Grounding answers in the merchant’s own products, orders and support history, so the feature is specific instead of generically plausible.
- 03
Human in the loop where it writes
Preview, approve, undo. Anything that edits products, tags or published content is reviewable before it lands and reversible after it does.
- 04
Vision features on the storefront
Virtual try-on and image-driven experiences on product pages, with the latency and image-cost budget treated as a requirement.
- 05
Honest limits
Where the model is unreliable, the feature says so rather than guessing. The fastest way to lose a merchant’s trust is a confident wrong answer applied to 4,000 products.
Where we have already done this.
Our own catalogue, not a case study written for this page. Every name below is public software you can open right now.
AI SEO Pilot
Scans a store and fixes the SEO issues it finds — bulk automated edits with review, which is the hardest version of this problem.
AI Easy Automation & AI Auto Post
Rules that act on orders, products and customers, and AI that turns a product into posts across social channels.
TryOnIA
AI virtual try-on on every product, so shoppers see fit before checkout — vision in production on a storefront.
What people ask before starting one of these.
What does an AI feature cost to run?
It depends on tokens or images per action, and it is measurable before launch: we model per-action cost, put a ceiling on it per tenant, and design the pricing around that. Features whose unit economics do not work get flagged as such rather than shipped and discovered later.
Whose model do you use?
Whichever fits the task and the budget — we work with Claude and OpenAI models, and route different steps to different tiers where that saves money without costing quality. The integration is written so the provider can be swapped rather than baked into every call site.
What about merchant data privacy?
Merchant data is sent only where the feature requires it, retention is set explicitly rather than left at a default, and the disclosure your listing needs is written as part of the work. Anything less is a compliance problem waiting for App Review.
More general questions — cost, ownership, what happens after launch — are answered on the home page FAQ.
Have a ai product features project?
Send the spec and an engineer answers within 48 hours — how we would build it, the rules that apply, and a price. The first working build follows days later, not quarters.
- A costed answer within 48 hours
- A working build you can click in days, not quarters
- Fixed price or T&M, quoted before we start
- Server, domain, SSL and monitoring handled — or shipped into your cloud
- You own the repo, the Partner account and the infrastructure
Prefer the company site? devcloudsoftware.com