Tech · case study
The AI lead pipeline
A supervisor types one line in Gujarati. The CRM gets a lead.
Bull's field supervisors talk to farmers all day and report in Telegram groups, in Gujarati, Hindi and Hinglish, in whatever words come naturally: a name, a phone number, a village, how many bags, when he wants to sell. Those messages were the company's lead flow, and they scrolled away unread.
The pipeline listens to all twenty groups and turns each message into a structured lead in the CRM, without asking anyone to change how they write.
Before
- Leads existed only as chat messages in twenty groups.
- Nothing reached the telecallers unless someone retyped it.
- Edited messages and repeat farmers became duplicates.
Now
- Every message is parsed into name, phone, village, crop, quantity and intent.
- A fifteen-second edit buffer catches corrections before they become a second lead; phone-number matching catches repeat farmers.
- Leads are tagged for urgency and size, so telecallers know whom to ring first.
- Batches go to the CRM on a schedule that respects its rate limits, with no messages lost.
How it works
Listen
A persistent Telegram session follows all twenty centre groups in real time.
Wait for the editCheck
Messages are held briefly, because supervisors correct a name or a number seconds after sending.
Parse
Gemini reads the line and returns the structured fields, in whatever mix of script and language it came in.
Deduplicate
A farmer who already exists is updated, not created twice.
Push
Leads go to the CRM in batches with their tags; telecallers see them on their list within the hour.
What it changed
- Four hundred-plus leads reach the telecallers every day with nobody typing them in.
- Parsing holds at around ninety-nine percent on real messages from every centre.
- The ground team kept writing exactly as they always had.
Built with Python with Telethon, Gemini 2.5 Flash, SQLite, FastAPI with a WebSocket dashboard, on AWS EC2.