01 / AI platforms
Dailoq — An AI customer-messaging assistant built for Azerbaijan
One assistant across Telegram, Instagram, WhatsApp, Messenger and the web — it answers from your documents, takes actions and hands hard questions to a human.
- 5channels — Telegram, Instagram, WhatsApp, Messenger, web
- ~98%of answers correct
- 15+companies using it
01 / Overview
Dailoq is an AI assistant for customer conversations, built for businesses in Azerbaijan. Whichever channel a customer writes from — Telegram, Instagram, WhatsApp, Messenger or a website widget — the same assistant answers, in Azerbaijani, Russian or English. It works from the company's own documents, carries out simple tasks such as booking a table, and passes difficult conversations to a human operator with a summary.
02 / The problem
Small and mid-sized businesses answer the same questions in several messengers by hand, often out of hours and in a mix of Azerbaijani and Russian. Global chatbot platforms don't understand local dialects, letter substitutions or mixed-language messages, price in foreign currency and tend to invent an answer when they don't know one.
03 / Our approach
We built the assistant around the local language first: Azerbaijani — with dialect phrases, letter substitutions and mixed AZ–RU sentences — is tested on its own. Answers are retrieved from the client's files and every reply shows its source; when the assistant can't confirm something it says so and offers a human. Every change to the bot's brain has to pass an evaluation set before it reaches customers.
04 / The result
More than 15 companies use Dailoq. They get one assistant for every channel that answers from their own documents — around 98% of answers are correct, with at most 2% errors — understands voice messages and photos, books and confirms actions and hands over to operators with full context, all managed from a single no-code panel with conversation history and monthly reports.
05 / Key features
What it does.
- 01
One brain, every channel
Telegram, Instagram, WhatsApp, Messenger and a web widget all work, the Meta channels through Meta's official integration. The conversation carries on when a customer switches channels.
- 02
Native Azerbaijani
Understands dialect, letter substitutions and mixed AZ–RU–EN messages, and replies in the customer's language.
- 03
Answers with sources
Searches PDF, DOCX, XLSX, TXT, Markdown, URLs and FAQs by meaning; every answer shows the file it came from, and unanswered questions become FAQ entries in one click.
- 04
Voice and images
The same multimodal model understands voice messages and photos — a menu, for example — and answers from them, with no separate OCR step.
- 05
Takes actions
Asks for confirmation, then books a table or calls the client's own system through a REST connector — with an idempotency key, so nothing runs twice.
- 06
Human handoff and panel
Hard questions go to an operator with a short summary; owners manage tone, knowledge, channels, team roles and history from one no-code panel.
06 / Architecture
How it is built.
Evaluation gate
A growing set of real questions — about half in Azerbaijani, 30% Russian, 20% English — runs on every change; if the score drops, the release is blocked.
Tenant isolation
Role-based access plus database-level isolation keep each client's data apart; connection keys are stored encrypted and an append-only audit log records every action. Production runs on a Contabo server in Europe behind Cloudflare.
Model routing
Adapters for several AI models, and a routing layer that picks the model for each request — switching provider is a configuration change, not a rewrite. Embeddings come from a self-hosted bge-m3 model, and semantic search runs inside PostgreSQL with pgvector.
07 / Technology
Tech stack.
- Backend
- C#
- ASP.NET Core 8 (.NET 8)
- Frontend
- Next.js (App Router)
- React
- WebGL 3D mascot (GLB)
- Data
- PostgreSQL 16
- pgvector
- Redis (cache and queues)
- AI
- Multi-model adapters with per-request routing
- Multimodal model for voice and images
- bge-m3 embeddings (self-hosted)
- Semantic search / RAG
- Channels
- WhatsApp, Instagram and Messenger (Meta)
- Telegram Bot API
- Web widget (13 KB, streaming)
- Infrastructure
- Contabo (Europe)
- Cloudflare
- Security
- Multi-tenant isolation
- Role-based access
- Encrypted credentials
- Audit log
