Event chatbot for Kortrijk Xpo
Academic team project for an external client, 2025
- My role
- Lead developer in a team of four: scraper, .NET API, retrieval, analytics backend, Docker
- Stack
- .NET, C#, Azure OpenAI, Cosmos DB, Next.js, Python, Scrapy
- Context
- Team project for Kortrijk Xpo
A chatbot for three Kortrijk Xpo trade fairs: Abiss, Artisan, and Flanders Flooring Days. It answers visitor questions from the event websites' content, links to registration, and records analytics for the organisers.
Problem
Answers from content that keeps changing
Visitors wanted quick answers about exhibitors, stands, and event details, and that information was spread over several event websites. The system had to collect the content, make it searchable by meaning, base its answers on what it found, and be configurable for each event.
Architecture
From ingestion to chat and analytics
Contribution
From source website to answer
- Started the project and wrote most of its code: 97 of the 124 commits.
- Wrote the Python pipeline that scrapes and cleans the event websites and uploads the result.
- Wrote nearly all of the .NET API: embeddings, the similarity search, chat, JWT authentication, registration tracking, and analytics.
- Connected Azure OpenAI and Cosmos DB, including retries for bulk uploads that hit rate limits.
- Added a prompt per event, conversation memory, stand-number lookup in several languages, a cache for context documents, input sanitization, and rate limiting.
- Put the backend, the pipeline, and the dashboard in Docker Compose, and built the dashboard's sign-in and chat history views. Teammates built most of the dashboard charts.
Result
One backend for several events
One backend serves all three events, each with its own Cosmos DB container, prompt, and chat widget. The search loads every embedded document and ranks them in memory, which is fast enough for a few event websites but would need a vector index to grow much further.