An AI email assistant that turns car parts enquiries into supplier requests and customer replies, drafted for the team to approve.
AGL Assistant
01 / Overview
A parts business that runs on email
Auto Group Logistic sells car parts to customers in Slovakia, Czechia and Croatia. Questions about prices, availability, orders and complaints arrive by email, and many can only be answered once a supplier has replied.
Created at BeCode, the assistant runs on a NestJS API and background worker with PostgreSQL and Redis, a Next.js interface for the team and OpenAI models that read, sort and draft the mail. As part of the team, I worked across the AI pipeline, from supplier routing to learning how each mailbox writes.
02 / Challenge
Two conversations behind every question
A price or availability question turns into two conversations, one with the customer and one with a supplier. A supplier must never learn who the customer is, so nothing could simply be forwarded. Part and vehicle details were rewritten by hand, and vague first messages had to be chased for the missing part or car.
Suppliers lived in separate Slovak, Czech and Croatian spreadsheets, each with its own contacts and languages. Choosing whom to ask and writing to them in their language was repeated for enquiry after enquiry.
03 / Approach
The model drafts, the team decides
A background worker picks up new mail from every connected mailbox. OpenAI models sort each enquiry into one of eight request types, check that it holds enough to act on and read part numbers even from attached screenshots and PDFs. Vague messages get a reply asking for what is missing. Complete ones become a request to a supplier chosen from one shared database, written in the supplier's language without naming the customer. Supplier replies are read for price, stock and delivery time. Drafts then wait for approval, written in a tone learned from the team's own sent mail.
04 / Results
An assistant that grades its own drafts
Each draft is now compared with the reply a colleague actually sent and given a score, so only the weak ones reach a reviewer. Their corrections can then shape the drafts that follow.
08
Request types recognised
Every customer email is classified, and each type has its own rule for what makes it complete and where it goes next.
06
Languages, one inbox
Drafts are written in Slovak, Czech, Croatian, Slovenian, German or English to match each customer and supplier.
~9,900
Sent emails imported
The mailbox's own sent history sets the tone of new drafts and supplies the replies each draft is scored against.
