InvoiceBridge

Inspiration

The main inspiration behind this project was to help customers of bunq. This goal was central to all of our brainstorming and the ultimate project we decided upon.

What it does

Our project is an Agentic AI solution that automates the process of linking invoices received via email to transactions in bunq. Invoices contain rich data, both itemised and textual and associating this with transactions in bunq enriches the data and context for customers to have a better overview in a single platform. We also see subsequent integrations with tools like Finn more powerful because there is greater context with which Finn can operate.

How we built it

Our stack consists of:

Python
CrewAI - our multi-agent framework
Combination of Llama Nemetron and OpenAI (for structured outputs)
Gmail integration used to develop tools for searching and extracting from emails/attachments

Challenges we ran into

With CrewAI, we have defined Pydantic models that define the outputs of various tools/stages/agents. We could not find models on the NVIDIA playground that supports structured outputs, which made it difficult for us
Getting a bunq account with the transactions that we needed to test this also proved to be difficult.

Accomplishments that we're proud of

We came up with an idea that we genuinely believe in and one that solves a problem for customers of bunq. Feedback from various bunq employees who have asked us have been overwhelmingly positive
We have implemented a POC using tools and technologies that are new to all of us. A great achievment for a 24 hour period.

What we learned

We learned a lot about implementing Agentic AI solutions and the unique challenges that come with that
We also learned a lot about bunq culture, that was reflected in one way or another across all the colleagues we spoke to