WhatsApp enquiry automation

The problem

Teaching Circle, a physics tutoring academy, was answering every WhatsApp enquiry by hand: timings, fees, payment confirmations, how to enrol. At more than 100 messages a day that took two to three people, and replies slowed down exactly when demand peaked.

Data

Two inputs: free-text WhatsApp messages, which are short and informal, and structured form submissions from students registering interest. A form submission arrives by webhook and starts the conversation.

Architecture

1. Webhook ingestionWhatsApp messages and form submissions arrive by webhook and are queued for workers.
2. Classify the enquiryCourse details, payments or enrolment, so each type gets the right handling.
3. Draft the replyLangChain orchestrates OpenAI's GPT-4 and GPT-5.1 mini to answer from the academy's own information.
4. Workers per conversationWorkers scale out while keeping each student's session separate, so concurrent chats never cross.
5. Log everythingConversations are logged to PostgreSQL for review.

Results

Results in production

Measure	Result
Enquiries handled	100+ a day
Classification accuracy	95%+ across 100+ tests
Manual response work	down 80%
Support staff needed	2–3 people replaced
Recorded outages	none

Where it breaks

• The 95% figure comes from the automated test suite, not from a labelled sample of live traffic. Real messages drift, and the suite has to keep up with them.
• One enquiry in twenty misclassified matters most for payments, where a wrong answer costs trust.
• Answers are only as current as the course and fee information the assistant is given. When the academy changes a fee, that source has to change the same day.
• WhatsApp only allows free-form replies within 24 hours of a student's last message. Anything later needs a pre-approved template.