LEO Chat — a store chatbot that shows the revenue

LEO Chat is LIONEX's own SaaS product. A chatbot platform for Ukrainian online stores: the bot answers questions from a knowledge base, shows products, completes orders with Nova Poshta delivery and payment right in the chat, and in the dashboard you can see which orders came after which dialogue, scenario or broadcast. We have kept it in production since July 2026 and use it ourselves: the widget in the corner of this page is LEO Chat.

The problem it solves

A store receives questions in messengers and on the site around the clock, and answers during working hours. Broadcast services and bot builders count subscribers and clicks, but do not show whether an order came out of it. LEO Chat closes the chain: question → answer or product → order in the chat → payment → a line in the revenue report tied to the scenario.

The bot does not create demand and does not guarantee sales. It removes a technical loss: questions left unanswered outside working hours, orders that never reached payment, broadcasts with no report in hryvnias. How much of that becomes orders depends on the assortment, the prices and how fast the managers are on complex questions.

How it works

1. Connection. A Telegram bot through the Bot API, or a widget on the site with a single script. The onboarding wizard walks you through channel → catalogue → bot template.
2. Knowledge and catalogue. A knowledge base from PDF/DOCX/TXT files, site pages and an FAQ. The catalogue comes from a Horoshop, Prom.ua or CSV feed on a schedule, or from a scanner that reads the product microdata on the site and builds a category tree.
3. Scenarios. A visual builder: messages, conditions, delays, actions, HTTP requests, handover to an operator, an AI answer, a product card, checkout, payment, a form. Validation before publishing, testing in an emulator.
4. Dialogue and orders. The AI answers within the knowledge base, remembers the context of the dialogue and suggests products with a “Buy” button. Cart → name → phone → city and Nova Poshta branch (the real directory), courier or self-pickup → cash on delivery or card → confirmation and the waybill into the chat.
5. Report. The “Revenue through the bot” dashboard: orders and amounts by channel, scenario and broadcast, the dialogue → order conversion, the scenario funnel by node, the share of dialogues closed without an operator.

What is inside

• Channels: Telegram and the web widget (a greeting with buttons, an inline request form, voice input, product cards, a cart). Instagram and Viber are planned.
• An AI consultant on the knowledge base with links to the sources; anything unknown is handed to an operator; memory about the customer; a list of “questions without an answer” to extend the base.
• A real-time Team Inbox: all dialogues in one window, handover bot ↔ operator, a customer card with orders and requests, the owner / manager / operator roles.
• Broadcasts by segment (tags, attributes, behaviour) respecting the Telegram limits, with the campaign revenue in the report.
• The Nova Poshta module: branches, parcel lockers, courier to the door, waybill creation, label printing, automatic status sync.
• Payment: cash on delivery, Monobank acquiring, LiqPay on the client's keys. Promo codes, refunds, an abandoned cart trigger.
• Forms and briefs: a step-by-step wizard in the widget, a conversational mode in Telegram; source labels (UTM, page) on every request.
• Analytics: attribution of orders to a scenario, an end-to-end funnel site → dialogue → request → deal, site web analytics with a heatmap and no personal data.
• For studios: several workspaces under one login, an agency plan with a shared pool of limits and a single invoice.
• Security: 2FA, session management, Google sign-in, API keys, a public knowledge base API, webhooks, an audit log. Data on servers in the EU; requests to OpenAI are not stored on the provider's side.

The revenue report is an attribution model (last-touch within a configured window), not accounting. It shows that an order was paid after a dialogue or a broadcast, and it does not claim that the order would not have happened without the bot.

Technically

The whole stack is TypeScript: one Fastify backend (REST API, receiving and sending messages, the scenario engine, WebSocket, the AI gateway), PostgreSQL + Prisma, Redis, NATS JetStream as a durable message queue, ClickHouse for events and reports, BullMQ for delayed sending. The dashboard is Next.js, the widget is Preact. The reliability layer: a transactional outbox, message deduplication, a dead-letter queue, payment reconciliation, a circuit breaker on every external API.

Our measurements. On 17.08.2026, on a test stack under a load of 300 messages per second, the engine was killed by force and NATS and Redis were restarted — not a single acknowledged message was lost or duplicated. On 18.08.2026 — 298 messages per second processed in real time; the integration suite ran 585 passing tests. On 04.07.2026, on a reference set of 45 questions, the AI gave 89.7 % accuracy and no answer outside the knowledge base, at an average cost of $0.0008 per answer. On 07.09.2026 the landing page responded in 0.21 s at 153 KB of HTML, and the widget script is 19 KB gzipped. This is the state of the product on the dates of the measurements.

Try it

LEO Chat is a self-service product with plans and a trial period without a card on its own site. The live demo is the bot in the corner of this page. If you need to connect a bot to your store, catalogue or CRM, leave the task with the agency: we reply within 2 hours and work under a contract.