Admissions Chatbot & Advisor Cockpit

The problem

Hướng Nghiệp Á Âu is an academy with 19 branches. Its admissions advisors were buried in messages from prospective students, and any chatbot answering for them had to meet one hard rule: never invent a course or a price.

What I built

I was the sole developer, freelance as AI Agent Consultant Engineer, from April to September 2026, and handed the system over at the end:

• A 24/7 admissions chatbot that answers prospective students on the academy's live CRM.
• An advisor cockpit — after each customer message, a background job drafts three replies and a lead summary for the advisor to pick from.
• Grounded answers — the CRM is the only source for prices and schedules; a course allow-list and topic pinning keep follow-up questions on the right course.
• Image reading through a self-hosted vision-language model, protected against SSRF.
• A provider-agnostic LLM layer, with routing between LLM providers set from a settings page.

How it works

Retrieval is hybrid RAG: pgvector similarity search combined with Vietnamese full-text search, so both meaning and exact course names are matched. Anything with a price or a date is read from the CRM at answer time rather than from embedded documents, which is what stops the bot from making numbers up. The allow-list rejects courses that do not exist, and topic pinning keeps a conversation tied to the course it started on. The whole system runs on-prem inside the academy's network, behind nine end-to-end test suites that had to pass before handover.

Outcome

Advisors accepted 97% of the drafted replies, so the cockpit took most of the typing off their hands while they stayed in control of what was sent. The system was handed over in September 2026.

Stack

• Hybrid RAG: pgvector + Vietnamese full-text search on PostgreSQL
• Provider-agnostic LLM layer
• Self-hosted VLM for images
• On-prem deployment inside the client's network