Kolloh Store: a commerce platform run from a chat window

The problem

A small merchant runs their business across five apps. Orders arrive in a chat thread, stock lives in their head, promotions are decided on instinct, and by the time they notice a regular has stopped ordering, that customer is buying somewhere else.

Store builders answer this with a dashboard. But a dashboard is somewhere you have to remember to go, and a merchant serving customers all day does not go there. The insight behind Kolloh is that the operations layer should come to the merchant, in the app already open on their phone, rather than waiting for them in a browser tab.

The agent, not the dashboard

A Claude-powered agent runs inside Telegram. New orders arrive in the chat with one-tap buttons to confirm, pack or mark ready, so an order moves from placed to delivered without a dashboard being opened. It writes a sales digest at sunrise and again at close of business.

It also watches for things a busy merchant will not: a regular who has gone quiet gets flagged with a discount offer already drafted, a new review arrives with a suggested reply to approve or edit, and stock runs low with a warning before the sell-out rather than an apology after it.

The dashboard still exists for everything that genuinely needs a bigger screen. It simply stopped being the only way in.

The machine-learning layer

A separate Python service, built on FastAPI, handles the work language models are the wrong tool for. Prophet produces a fourteen-day demand forecast with a confidence range rather than a single guess. Anomaly detection flags sales days that do not add up, calibrated to local weekends and holidays rather than a Western calendar. Customer lifetime value is scored with scikit-learn, segmenting buyers into champions, regulars and at-risk.

It runs as its own authenticated service rather than inside the main API, so a slow model run cannot block a customer checkout and the data science can be iterated on without redeploying the storefront.

Content generation that respects the market

Merchants sell on social platforms, so the agent produces the assets for it. A product photo becomes a branded banner sized correctly for Instagram square, story, Facebook or TikTok, with the caption, hook, price and hashtags written to match. A sale automatically picks up the discount badge, struck-through price and promo code.

Captions generate in Arabic, French or English with right-to-left handled properly — not a translated afterthought. A phone photo taken on the shop floor comes back with the background replaced by a clean studio scene.

Built multi-country from day one

Country is the tenant boundary. Currency, tax rules, categories, document requirements and available payment, delivery and SMS providers are all configured per market, so launching in a second country is an administrative task rather than a fork.

Underneath sit nine applications: a Fastify-based NestJS API, four Next.js front ends, background workers, the Telegram agent and the ML service. Governance runs on a six-role permission model — super admin, country admin, moderator, business owner, business staff and customer — where a country admin’s authority stops at their own market. Search runs on self-hosted Meilisearch, chosen after costing metered alternatives at the volumes expected.