OUR PRODUCT
Aisle
AN ASSISTANT THAT KNOWS WHERE YOU ARE
Address · Inventory · Slot · Logistics Engine
Most support chatbots read documents. For a business that delivers to an address, that isn't enough: the answer to every question changes with where the customer is.
“Can you deliver to me?” depends on whose service area the address falls in. “How much?” depends on which of four price tiers applies to that channel. “When?” depends on the dealer's working hours, the slot table and courier load. None of it is written down anywhere — each one is computed on the spot.
THE APPROACH · FACTS FROM THE SYSTEM, LANGUAGE FROM THE MODEL
Aisle is deliberately hybrid. Anything that has to be exact — dealer eligibility, stock, price tier, delivery slot, order status — Claude fetches by calling the platform's live APIs. Retrieval is reserved for free-form knowledge. The model cannot guess a price, because it never sees one.
Answers are not free-form output; they render from a registered template catalogue. Every answer has a plain-text equivalent and an audit trail. If a question falls outside the catalogue, the assistant doesn't improvise — it hands over to a person.
THE FOUNDATION · WE ENGINEER THE SYSTEM UNDERNEATH
We didn't bolt an assistant onto software we don't know. We have been the engineering team behind Su İste, a water and beverage distributor in Istanbul, for the past year: backend, admin panel, iOS and Android apps, and the Getir, Trendyol GO, Yemeksepeti and Hepsiburada integrations. We know the code and the data model from the inside.
That matters more than it sounds. An assistant is only as good as its reach into the system that computes the answer. Ours calls 14 endpoints we designed for exactly this: does the address fall inside a dealer's service area, is it in stock at that dealer, which price applies on that channel, when is the first open delivery slot, where is the order. A vendor who doesn't know the platform from the inside either scrapes a website or settles for an approximation.
Su İste runs on it across seven sales channels, handling thousands of orders a month.
Aisle is an acronym: Address, Inventory, Slot, Logistics Engine. The first four letters are the four jobs listed in the box on the right — address matching, stock, delivery slot, courier dispatch. The fifth is the engine that runs them.
BUILT ON · CLOUDFLARE, END TO END
Workers · AI Gateway · Vectorize · D1 · Durable Objects
Designed multi-tenant: a separate publishable and secret key per customer, origin-restricted widgets, signed webhooks in both directions, isolation per customer at the database and vector-namespace level. The embeddable widget is Preact in a Shadow DOM, under 60KB gzipped, with a 5KB loader.
WHERE IT CAME FROM · 1,463 REAL CONVERSATIONS
We built this from the record, not from guesswork. We read 1,463 messages Su İste's customers actually wrote and drew a 22-intent taxonomy out of them: what the assistant can answer, and what it cannot, written down one by one.
Two findings changed the product. About a third of the questions were specific to one order, which forced an identity-verification flow into an otherwise anonymous widget. Another 18% weren't about water at all — people were stuck in the mobile app. So Aisle knows both apps screen by screen, and sends users to a deeplink instead of describing a button.
STATUS · PRE-LAUNCH
The integration layer is running in production and authenticating: 14 endpoints, HMAC-signed webhooks, signed context tokens, 17 feature tests green. Two things stand between us and launch: the conversation engine, and the evaluation pipeline — a reference set drawn from those 1,463 messages, enforced as a gate in CI.
Aisle works for any business with dealer territories and delivery times: water, grocery, pharmacy. If you draw a boundary on a map, Aisle knows it.