Inspiration
Global sourcing and wholesale logistics are often bottlenecked by fragmented, legacy supplier portals. We recognized a critical need for an automation tool that doesn't rely strictly on clean APIs but can interact natively with the unstructured web to eliminate tedious manual data transfer.
How we built it
Fetch runs on an asynchronous Python backend (FastAPI) powered by the Amazon Nova suite: • Amazon Nova Act: We used Nova Act to complete repetitive UI workflows directly in the browser. It handles the heavy lifting of navigating supplier sites without APIs, filtering search results, and extracting key unstructured data. We also used it to complete complex purchase flows autonomously and act as a UI-based fallback when Shopify's GraphQL API fails. • Amazon Nova Pro & Lite (Bedrock): These models handled all non-browser tasks, including pure reasoning, product ranking, search autocomplete, and text generation.
Challenges we ran into
Managing multiple concurrent headless browser sessions without blocking our FastAPI event loop required careful thread pool and semaphore orchestration. Additionally, we learned that giving an AI agent one massive prompt reduces reliability. We had to architect our workflow to break down complex UI tasks into a sequence of smaller, explicit commands passed to the model.
What we learned
We discovered the immense value of Human-in-the-Loop (HITL) architecture. By implementing Nova Act's human approval pattern, the system can pause at a decision point, capture a screenshot of the current state, and present it to a human reviewer. This allows Fetch to fully automate the tedious parts of the checkout flow while safely escalating to a supervisor for final payment approval.
Built With
- alembic
- amazon-bedrock
- amazon-nova-act
- amazon-nova-lite
- amazon-nova-pro
- amazon-selling-partner-api
- asyncpg
- aws-ecs
- aws-secrets-manager
- docker
- fastapi
- hiredis
- httpx
- passlib
- postgresql
- pydantic
- pytest
- python
- python-jose
- pytrends
- redis
- shopify-api
- sqlalchemy
- uvicorn

Log in or sign up for Devpost to join the conversation.