Customer emails, answered from your own documents.
Small local AI models read each email, find the answer in the firm's own price list and terms, and write a draft. When they are not sure, a person decides. Nothing leaves the computer.
–emails the machine handled
–handled wrong
–sent to a person
–right document found first
How it works
1 · Sort
A small decision model picks question, order, complaint or other, and says how sure it is.
2 · Find
A search model finds the three closest parts of the firm's documents.
3 · Write
A writer model drafts the reply from those parts only, in the customer's language.
4 · Check
The decision model checks every price, time and number in the draft against the sources.
Run it yourself
ollama pull qwen3-embedding:0.6b # plus the decider and writer, see README docker compose up # page on http://localhost:8080 python3 run.py emails-100.json # re-run all 100 emails, updates this page
The firm, people and emails are made up. Results on this page come from one real run on a 16 GB laptop.