Case study · AI & Telegram Bots

Voice-powered Bot: listens to realtor voice notes and fills CRM cards automatically.

Real estate agents waste hours copy-pasting listing details. We built a Telegram bot (Collector) where realtors drop photos and record brief voice notes right from the property. AI under the hood parses details, writes a high-converting description, and creates a draft record. Bonus: agents don't need CRM access, protecting your database.

ProjectReal Estate Agency Automation StackPython, aiogram, Gemini API RoleAI Engineer · Backend Developer Key FeatureVoice Recognition + Entity Extraction
Voice instead of typing
20+ distinct fields
0 database leak risks
(no DB access given)
100% compatibility with
complex legacy ID rules
01 / Context

Why classic CRMs kill sales speed

The core problem of every real estate agency: realtors hate working with CRMs. Their job is doing showing tours and negotiations, not sitting in an office copy-pasting square footage, floor numbers, and pricing into 20 inputs.

To post a single listing to the website, the agent had to snap photos, head back to the office/home, copy photos to a PC, write SEO-friendly text, and manually fill the database. If a property was sent by a third-party agent, they had to re-type everything. It was a massive time sink.

The second pain point was security. Giving database (CMS) access to all freelance agents created a high risk of someone scraping the entire database right before resigning.

No Bot / With Collector Bot

— Before
  • Manual typing from keyboard
  • Agents waste hours formatting data sheets
  • Security leak risk (everyone has CMS access)
  • Agent writes copy & SEO keywords from scratch
  • Human errors when typing complex legacy listing IDs
+ After
  • Realtor sends voice notes and photos on the spot
  • AI extracts entities and fills database fields
  • Write-only access: data leaks are structurally impossible
  • AI generates engaging property descriptions
  • Bot handles complex legacy ID rules automatically
02 / Under the hood

AI working as a virtual assistant

The bot workflow matches natural human chatting. The agent drops photos and records a brief voice message (or forwards a text listing from another realtor). Then the automation takes over:

  • Transcription Voice Recognition. The bot converts the voice message into clean text. The raw audio transcription is also saved to the database "for history" to track original listing parameters.
  • AI Parsing Entity Extraction. An LLM analyzes the text and pulls exact metrics: price, area, floor, rooms, and district. These values are automatically mapped to the correct columns in Directus CMS.
  • Copywriting Description Generation. Based on raw metrics, the AI drafts a polished, engaging description for the public website catalog.
  • Legacy ID Format Formatting. The agency used a custom non-sequential ID system (e.g., an "X" at the end denotes "direct deal"). The bot handles these rules and maps IDs cleanly.

Zero Trust: Smart Security

Since the bot acts strictly as a **collector**, we can securely share its link with all freelance agents. They have no login credentials for Directus. They *cannot* view listings of other realtors or export any database tables.

Realtors simply feed the funnel. The bot builds a Draft, and the admin approves publication to the site in one click. Property processing time was cut from hours to seconds.

Need similar workflow automation? Tell me about your task — I’ll draft what is feasible.

Message on Telegram
Back to cases