Case study
Quick facts
- Role: Backend Developer
- Timeframe: Delivered MVP + deployment + initial support
- Platform: Telegram bot + Django admin + CRM integration
- Status: Released
- Team: Solo
Summary
Telegram chatbot for a travel agency that captures trip preferences from either free-form text/voice or a structured survey, then submits a complete lead to Bitrix24. Voice messages are transcribed with Yandex SpeechKit and confirmed by the user before submission. Bot content and survey configuration are editable via Django admin. Key highlights:
- Two intake modes: audio/text request or multi-step survey
- Voice-to-text with confirmation loop to prevent bad leads
- Admin-controlled questions, answers, texts, and contact methods
Problem
- Leads were collected manually and inconsistently. Missing details caused back-and-forth with customers.
- Needed voice support, but STT errors can corrupt intent and waste agent time.
- Survey content had to be editable without code changes and deployable on the client’s server.
Solution
I built a state-driven Telegram bot that guides users through either free-form intent capture or a controlled survey, then collects contact method + phone number with validation. Once all required fields are present, the bot assembles a single payload and pushes it to Bitrix24 via API. Voice recognition runs asynchronously and is toggleable from Django admin.
- End-to-end flow: intake → validation → CRM submission
Architecture
- Django app with webhook endpoint for Telegram updates
- Finite-state conversation engine (step / sub-step / micro-step stored per user)
- Survey engine driven by DB models (questions, ordered answers, “own option” branches)
- Voice pipeline: Telegram file download → STT (short vs long) → user confirmation
- Celery + RabbitMQ for background tasks (STT, CRM posting)
- Django admin as CMS for bot texts, keyboards, contact types, survey content
- Dockerized deployment with Postgres, worker, and web service
Hard problems solved
- Built a reliable state machine that survives restarts and handles re-tries without losing user context
- Implemented voice STT with a confirmation loop to protect lead quality
- Supported long-audio transcription via object storage + long-running STT polling
- Prevented invalid CRM payloads by enforcing strict phone validation and required-field completion
- Designed survey data model to support ordering, grid keyboards, and “custom answer” branches
- Made bot behavior configurable via admin (voice mode toggle, editable texts and options) without redeploys
- Structured async CRM delivery (sync in dev, background in prod) to keep chat UX responsive
Impact / Results
- Standardized lead intake and ensured each lead reaches Bitrix24 with complete, structured data
- Reduced ambiguity by confirming STT results and enforcing validation before submission
- Enabled non-technical staff to update survey/questions/texts through admin
Tech stack
- iOS:
- Architecture:
- Backend/Infra: Python, Django, Celery, RabbitMQ, PostgreSQL, Bitrix24 API, Yandex SpeechKit, Yandex Object Storage (S3), Docker
- Tooling: Django Admin, telebot, adminsortable2, gunicorn, nginx
