AI Support Automation & Knowledge Retrieval
Helping support staff read less and answer better.
- Role
- Sole developer
- Status
- Partly in production
- Stack
- Python · Django · OpenAI & Anthropic APIs · Embeddings · RAG
Suggested replies are in development and not yet deployed.
How it fits together
01
Context
Support staff spent time sorting tickets and reading long threads before they could respond.
I built AI workflows that categorize tickets and summarize threads. I’m now developing reply suggestions grounded in past support conversations.
02
Constraints
- Support conversations can contain personal data, which shapes what can be indexed and how.
- Suggestions must stay suggestions: a person reviews every reply.
- Output has to be grounded in real past answers, not invented.
03
What I built
- Ticket classification in production, using OpenAI and Anthropic models inside Django workflows.
- Thread summarization in production.
- Prompt versions compared against labelled test cases, not judged by eye.
- In development: retrieval over the support archive to ground suggested replies, with a person reviewing every reply.
04
Tradeoff: Assist, don’t automate replies
Sending replies automatically would be faster, but support answers carry the business’s voice and its commitments. Keeping a person in the loop costs time and makes the system trustworthy.
05
Outcome
Support staff get categorized tickets and thread summaries; reply suggestions are being built next.