MeerBot

Problem

Businesses want AI chatbots that answer from their own knowledge base, not generic GPT responses, and that managers can take over without switching tools. Existing options are either too expensive (enterprise), too limited (simple GPT wrappers without a knowledge base or human handoff), or require an engineering team to deploy. There was no affordable platform that combines a web cabinet, Telegram bot, website widget, multichannel inbox and a multi-model AI under one roof.

Solution

Built a two-part platform: AI Core (self-hosted LLM engine with OpenAI-compatible API) and Platform (SaaS cabinet + Telegram bot + website widget). Stack: Next.js 16 + TypeScript, Prisma/PostgreSQL, pgvector for 1536-dim embeddings with cosine similarity, Grammy for Telegram with Redis-backed sessions, LiteLLM proxy in front of GPT-4o / Claude / Gemini. Knowledge base pipeline: upload PDF / DOCX / URL, structure-aware DOM walker, semantic chunker, embeddings via text-embedding-3-small, top-K retrieval per query. Manager tools: multichannel inbox (Telegram + website unified), Discord-style replies and message editing, quick replies with media attachments, customer profile with devices/subscriptions/payments, manager notes injected into AI context, auto-escalation to humans on LLM failure. AI Copilot 2.0 in 3 tiers (fast / standard / powerful) with per-conversation persistence and daily Telegram digests. Auto-Improver: AI proposes instruction improvements from real dialogs. Problem Analytics with LLM-driven insights. Public API gateway v2 with vision, Bearer auth, OAuth, rate-limiting, AES-256-GCM encryption and SSRF validator. YooKassa + 54-FZ receipts, per-token billing with 1.4x markup, hard caps and usage CSV exports.

Results

• +In production with 2 paying clients (e-commerce, support)
• +Multichannel inbox: Telegram + website widget unified for managers
• +Public OpenAI-compatible API (v1) with Bearer/OAuth and vision support
• +Multi-model AI: GPT-4o, Claude, Gemini via LiteLLM proxy
• +RAG pipeline: from document upload to vector search in seconds
• +AI Copilot in cabinet: 3 model tiers, daily Telegram digest, per-conversation memory
• +Auto-escalation to human managers on LLM failure
• +Security hardened: JWT allowlist, HMAC, CSP, AES-256-GCM, SSRF validator

Role & timeline

Co-founder & developer. Started with a free demo bot to validate the idea. Building AI Core, Platform, Telegram bot, website widget and public API with a partner