products in one bot
Tours, excursions, eSIM, FX, insurance plus seven more.
← Back to all cases
Travel · Telegram
Travel bot in Telegram, a CRM admin and an AI lead-gen pipe from public chats
A bot for people who holiday in Thailand: tour picker, excursions, eSIM, currency exchange, insurance, bonuses and giveaways. The business gets a separate CRM admin to manage content, segments, broadcasts and the partner program. Plus a lead-gen channel: a parser listens to public TG chats and a neural net picks travel questions and turns them into prospects.
01screen
02screen
03screen
The pain
- A customer juggles a website, Instagram, messengers and Google Sheets to buy a tour.
- Leads slip through: a blogger brought a customer — who paid and through which channel? Nobody remembers.
- Newsletter subscriptions are all-or-nothing: no segmentation by interest.
- Bot content needs developer attention for every change: each message, photo or video — a separate ticket.
- Giveaways are run by hand: participant lists, links, winner announcements.
- Other travel agencies wait for the customer to DM them. What if we actively look for customers on Telegram ourselves?
What we did
- A Telegram bot with 12+ products: tours, excursions, eSIM, FX, insurance, hotels, transfers, free guides, giveaways.
- Tour and excursion pickers as step-by-step flows: in a few taps the customer gets a personalised quote from the operator.
- A separate CRM admin on a web UI (Next.js + React) for everything inside the bot: content editing, message templates with media, broadcasts by segment, user labels, send log, stats, giveaways and partner program — without a developer.
- User segmentation via a labels system: broadcast only to people who looked at Phuket, or only to past buyers.
- Partner program: unique blogger links, automatic sales attribution and payout calculation without Excel.
- A giveaway builder in the admin: spin up an active giveaway, the bot collects entries and records winners.
- An idea we proposed ourselves: a separate lead-gen channel that parses public Telegram chats and runs the questions through an LLM. The parser listens live, the AI tags «question from a likely traveller», and the operator gets a ready lead with context.
—
What we ended up with
admin without a dev
Content, broadcasts, segments, giveaways, partners — all by the operator.
operator-free lead capture
The bot, broadcasts and AI parser work around the clock.
leads from public chats
AI finds travel questions in open TG chats and surfaces them to the operator.
—
What we built
- 01
Travel bot in Telegram
12+ products in one bot — tours, excursions, eSIM, FX, insurance, hotels, transfers, guides, giveaways.
- 02
Tour and excursion pickers
Step-by-step resort and trip-format selection, personalised quote from the operator.
- 03
CRM admin
Bot content, message templates with media, broadcasts by segment, user labels, send log, giveaways and partner program — all through the UI without a developer.
- 04
Partner program
Unique blogger links, automatic sales attribution and payout calculation.
- 05
Segmented broadcasts
User labels (Phuket, Pattaya, buyers) — surgical broadcasts.
- 06
Giveaways
An active-giveaway builder, automatic entry collection and winner recording.
- 07
AI TG-chat parser
Listens to public chats, AI picks travel questions, operator works with a ready lead.
— a similar project for you
Bot as a sales channel — and a lead pipe from TG chats?
We’ll build a bot with a CRM admin and a Telegram lead-gen channel for your business. Tell us the problem — we come back with an architecture and an estimate within a couple of days.
← Back to all cases
Telegram bot · CRM admin · AI lead-gen
Travel bot, a CRM admin and an MTProto parser of public chats with LLM-based classification
Three connected products in one system: a travel bot on node-telegram-bot-api (polling), a CRM admin for content and operations, and a real-time MTProto parser of public TG chats on gramjs with a background LLM worker (Gemini 2.0-flash, plus a stand-by local Qwen 2.5 1.5B on FastAPI). PostgreSQL holds a 13-table schema — from users and sales to segments, templates and logs. The AI lead-gen channel was an idea we proposed ourselves and baked into the architecture together with the bot.
01screen
02screen
03screen
—
Stack
- Bot
- Node.js · TypeScript 5 · node-telegram-bot-api v0.64 · polling · ESM
- CRM admin
- A separate web app: Next.js · React · TypeScript · ADMIN_IDS access · action logs
- Parser
- gramjs (MTProto, telegram v2.26) · live NewMessage listener across hundreds of chats · events arrive over MTProto push, not polling
- AI
- Gemini 2.0-flash for question classification · stand-by local LLM service (FastAPI + Qwen 2.5 1.5B-Instruct via llama-cpp)
- Workers
- A Node worker in its own container · 60-second cycle · updates parsed.category
- Data
- PostgreSQL 15 · 13 tables · 25 indexes · 9 updated_at triggers · JSONB for media templates
- Integrations
- Google Apps Script Web App (parser chat sources) · Google Sheets (CRM exports)
- Infra
- Docker Compose · 5 services (postgres / bot / parser / worker / llm) · one container per role
—
Architectural decisions
- 01
CRM admin as a separate web app
A dedicated Next.js + React web UI: bot content management, broadcasts, user labels, giveaways and the partner program.
- 02
Custom MessageFlowService instead of FSM scenes
A polling bot with no server-side sessions: instead of Telegraf scenes we track «message blocks» (primary + related message_ids), clean them up on navigation, and mark persistent messages that survive a return to the menu.
- 03
An MTProto parser on gramjs
A listener subscribes to UpdateNewChannelMessage / UpdateNewMessage, matches chats by chatId / username / title, filters texts that contain «?», extracts authors via InputUserFromMessage and writes into parsed(chat_id, message_id, text).
- 04
Source throttling
Parser chat sources live in an external Web App; the parser refreshes them at most once every 10 minutes per chat — to avoid Telegram FloodWait.
- 05
Two-stage AI pipeline
First the parsed table is populated by the parser. Every 60 seconds categoryWorker grabs an unprocessed row, sends it to Gemini 2.0-flash with a Russian prompt («pharmacy, tours, excursions, FX, other») and saves category + model into parsed.
- 06
Two-model LLM strategy
Gemini in the cloud for classification quality, plus a stand-by local Qwen 2.5 1.5B-Instruct via llama-cpp (FastAPI with an OpenAI-compatible /v1/chat/completions). Switchable without touching the worker.
- 07
DB-level partner program
users.referral_code + bloggers with a unique slug, sales.referred_by → bloggers_leads with automatic payout calculation. No manual Excel rolls.
- 08
Labels as segmentation
user_labels(user_id, label) — composite PK that powers precise broadcasts: «everyone who looked at Phuket», «March buyers», «inactive 30 days».
- 09
Message templates with media
message_templates.media_files holds a set of Telegram file_ids in JSONB; the bot resolves them without re-uploading — saves traffic and keeps one content edit point.
- 10
Roles isolated in Docker Compose
postgres, bot, parser, worker and (optionally) llm — each in its own container. The worker scales horizontally without restarting the bot.
—
Project metrics
12+products inside the bot
13tables in the schema
25indexes
5Docker services
2LLM providers in the architecture
10 minparser source throttling
60 sAI worker cycle
1 + 1developer + AI integrator
—
Inside the system
Polling bot on node-telegram-bot-api
No webhooks, no server sessions — state via MessageFlowService + chat-id maps.
CRM admin with content editing
A separate Next.js + React web UI. All bot copy, media and buttons editable through the UI — no release. Access via ADMIN_IDS.
Message templates with media
message_templates.media_files JSONB — Telegram file_id reuse for heavy content, a single edit point.
Multi-part broadcasts
Draft mode, media groups, label filters, send stats.
Label-based segmentation
user_labels with a composite PK, broadcasts by any combination of tags.
Partner program
bloggers, bloggers_leads, sales — automatic sales tracking and payouts with per-blogger overrides.
Giveaway builder
giveaways + giveaway_winners, slug-routed participation via callback_data.
Real-time MTProto parser
gramjs listens to public chats, filters questions, writes into the parsed table with chat_id and message_id.
AI question classification
The worker calls Gemini 2.0-flash and saves category + model into parsed for downstream lead selection.
Stand-by local LLM
A FastAPI service on Qwen 2.5 1.5B-Instruct, OpenAI-compatible — switchable without touching the worker.
Google Sheets integration
Parser chat sources and CRM exports go through a Google Apps Script Web App.
Logs in the DB
A logs table with level/context — no separate observability stack at launch.
— a similar project for you
Bot as a sales channel — and a lead pipe from TG chats?
We’ll build a bot with a CRM admin and a Telegram lead-gen channel for your business. Tell us the problem — we come back with an architecture and an estimate within a couple of days.