Messy inbox in. Real work out.
Python
Streamlit
SQLite
LLM API
Customer messages arrive unstructured — different wording, tone, and detail every time. Someone has to read each one, work out what the customer needs, judge how urgent it is, and decide what happens next, before any real work starts.
This tool does that first pass: it sorts the message, ranks its urgency, routes it to a department, and drafts a reply — then hands the whole thing to a person for a yes or no.
- The AI drafts, a human decides. Every field is editable and nothing is final until someone approves it. The model is explicitly barred from promising refunds, replacements, or delivery dates — those are business decisions, not classification tasks.
- Model output is never trusted directly. A validation layer checks every field against approved values before anything is stored, so an unexpected response becomes a safe default rather than a bad record.
- The AI provider is configuration, not code. Swapping the underlying model is an environment variable, not a rewrite.
