Channel K06 · 2026
Reply Engine
Live demo · built on my own initiative
Four AI agents recognise the question, search the knowledge base and write a draft.
Try live (in Dutch)
In short
- Built for
- Businesses with a lot of recurring customer questions
- The problem
- Looking up and answering the same questions every day, while the answer is already written down somewhere
- What it delivers
- A draft reply from your own knowledge base, checked by a quality check: a measured median of 3.6 seconds once the demo is running. Try it live with your own PDF.
Overview
Four AI agents work together on customer questions: they recognise the question, look up the answer in the knowledge base, write a draft and check it. Upload your own knowledge base as a PDF and ask it questions straight away.
Challenge
Businesses spend hours every day answering recurring customer questions. Looking up, wording and answering each question again, while the answer is already in the knowledge base.
What it does
- 01Recognise
What does the customer actually want to know?
- 02Search
The answer is looked up in the business's own knowledge base
- 03Draft
Based on what was found, a draft reply is written
- 04Check
A quality check looks at whether the answer is correct and complete
Result
- Answers from your own knowledge base, checked by a quality check
- A measured median of 3.6 seconds per draft reply, once the demo is awake
- You see every step happen live
- Upload your own PDF; it is deleted automatically after 1 hour
Technology
A pipeline of four specialised agents: intent classification, knowledge base retrieval via ChromaDB, answer generation with OpenAI and a quality check. SSE streaming, so the user sees every step live. FastAPI backend, deployed with Docker on Render. The four agents: intent, retrieval (semantic search in ChromaDB), draft and quality.