Chatbot answering visitor and customer enquiries in real time; trained, tested and refined on real user queries and feedback.
- My role
- AI Specialist — chatbot, evaluation loop
- Links
- GitHub (planned) ↗
- Stack
- PythonFastAPI
- LLM · RAG · Python · FastAPI
- Microsoft-ecosystem fit
▸ Copilot-style assistant pattern
▸ Responsible AI: evaluation on real queries
24/7
Live enquiry answers
[REPLACE]
Answer accuracy before → after tuning
Source · Level Up Training Center website (rebuild) and the “Ask Level Up” assistant captured answering in the Level Up build; personal details blurred
01 — Problem
Staff handled repetitive course and schedule questions during the day, and enquiries went unanswered after hours.
02 — Approach
- 01
Grounded answers in centre information: programmes, NCLDP packages, policies and contact details.
- 02
Inside the portal, the assistant answers from the signed-in user's own sessions, invoices and progress, never another family's.
- 03
Suggested questions cover the most frequent enquiries; real conversations are logged, misses labelled and prompts and retrieval refined against that set.
03 — Architecture
System architecture
S01
Visitor
S02
Chat widget
S03
Retriever
S04
LLM
S05
Feedback log
S06
Re-tune
Outcome
Enquiries are answered immediately, with accuracy improving through a structured feedback loop.
Next project
Road Health Tracker
