Case study
Paste an order, get a label: an AI shipping desk that quotes every carrier through ShipStation and checks out to a printable label.
Screenshots of the app running in demo mode, with simulated rates and fictional addresses.
Shipping one package usually means five tabs of copy-paste: lift the address out of an order email, key it into a carrier portal, compare prices across carriers, then buy and print the label somewhere else.
The whole desk collapses into one paste. The assistant extracts the sender, recipient, package and contents, quotes every carrier at once, and hands off to a checkout that is already filled in, so the path from order to printable label is a couple of clicks.
A Python + Streamlit shipping desk that started as a four-person team prototype and was rebuilt end to end. An AI assistant (Groq tool calling, with an optional AWS Bedrock Agent) reads a pasted order email, packing slip or address block, builds the shipment, and fetches exact rates from every connected carrier through the ShipStation API v2. Rates come back with carrier logos and Cheapest / Best value / Fastest picks; checkout opens pre-filled and ends in a real label PDF with a tracking number. Around the assistant sit Dashboard, Rates, Checkout, Shipments and Tracking pages.
The project started as Shipping Agent Assistant, a four-person team prototype. It was a single-page Streamlit chat app wired straight to a deployed AWS Bedrock Agent through boto3. The agent called the ShipStation API for rates and tracking, and answered in the chat as prose. Its source still lives in the original repo, and the rebuild keeps that Bedrock Agent as an optional assistant provider.
Screens are recreated from the legacy app's real UI with fictional demo data.
A multi-page Streamlit app (Dashboard, Rates, Checkout, Shipments, Tracking, Assistant) sits on a small shipping_agent package. A ShipStation API v2 client handles rates, estimates, labels, void, tracking and address validation; a Groq tool-calling agent (or an optional AWS Bedrock Agent) calls quote_rates, create_shipment, track_package, validate_address and list_carriers, and hands a finished shipment to a pre-filled checkout. With no keys, a demo client stands in so every page still works.
The problem
A chat model will happily invent a rate or say a label was bought when it was not.
What I did
Every price comes from a ShipStation tool call, the prompt forbids inventing prices or claiming a purchase, and the rate table and label panel render straight from API results.
The problem
A production key can buy real, billed labels, which is dangerous in a demo or a misconfigured deploy.
What I did
Labels are bought with the sandbox key by default; the production key only buys when ALLOW_LIVE_LABELS=1, and otherwise checkout hands payment to ShipStation or the carrier. The app never collects card numbers.
The problem
The cheapest rate is often USPS Media Mail, which is only allowed for books and media.
What I did
Media Mail is only picked when the contents are books or media; Cheapest and Fastest prefer services with a delivery estimate, and Best value ranks days saved per extra dollar.
The problem
The legacy build answered in prose from a Bedrock Agent, so rates could not be compared, picked or turned into a label.
What I did
Moved pricing, labels and tracking into typed ShipStation tools behind structured tables and checkout, and kept the Bedrock Agent as an optional provider rather than the core.
Legacy: Bedrock prototype
The team's first version: a Streamlit chat app calling a deployed AWS Bedrock Agent via boto3 with IAM-scoped invocation, plus quick-action buttons.
ShipStation rebuild
Exact rates, estimates, labels, void, tracking and address validation through the ShipStation API v2.
AI checkout
Groq tool-calling agent that builds shipments from pasted text and hands off to a pre-filled checkout.
Live tracking coverage
Tracking falls back to carrier pages or simulated scans when the plan or sandbox lacks API tracking.