A seemingly unsolvable problem

When it comes to our family's finances, it's always been a struggle to keep track of transactions cross multiple accounts, multiple cards, and multiple bank institutions. Not only that but monthly reports only come out, well, monthly. And no bank allows the export of the transactions available through your web portal until the month is over. This forces you to use the native tools the financial institution provides which are almost always falling short.

Additionally, the financial institutions that do allow you to create a budget with transaction auto-classification, live tracking, and notifications, do not allow sharing between partners of the same account. In 2026 do we really need to use Google Sheets to track our transactions, copy-pasting until the cows come home and having to play with formulas that break as soon as a dot appears instead of a comma. And don't start me on the interpretation of "1000" vs "1'000" that breaks formulas too!

We really needed to find a solution that would be future-proof, could be shared, be as live as possible, could auto-classify transactions, all the while making sure our financial data was not sold to brokers across the globe and stays ideally at home. As we will see, it blew out of proportions and changed our lives, but at that time those were our only requirements.

Why didn't I think of this before

It really feels like LLMs are everywhere these days and it seems like we are only scratching its surface of its capabilities. As a personal experience, their rise of performance and ease-of-use has been a tremendous help in both my career and with my late night tinkering projects.

Although it may be obvious to some, building an app with the help of AI to solve my family finance kerfuffle came like a lightning bolt. Why never did I think of this before? It felt like a dark corner of my mind was now emanating light.

I am using AI every day for work, but not using it to solve my own problems is a mistake I will only do once. Let's see where this will lead.

Using AI to make a plan for the AI to build so the AI can support it easily

I think this is one of the most underrated feature of LLMs. Given different instructions, it can act as the planner, or the implementer, or the reviewer, or the support technician, giving you different insights at each step. So I began with a simple prompt :

What would be the best language for you to build a simple web app with a database that would be hosted on a low-resources linux server? I want the app to be installed as a PWA on target machines and be responsive (Desktop and Mobile)

It asked me a couple follow-up questions and then settled on .NET + Blazor + SQLite. I then asked it to plan a simple back-end architecture that would be simple to support and edit in the future and to make sure to break up the front-end files to make sure to use as few tokens as possible when making edits. After these steps, it becomes as simple as asking for direct features. For example :

I want you to create a simple transaction tracker that would be divided in 3 pages : 
 - Page 1 : Dashboard - Shows simple donut graphs about data you would find relevant for a monthly report
 - Page 2 : Transactions - I want to be able to import transactions and classify them based on the categories configured in Page 3
 - Page 3 : Categories - I want to be able to set categories and assign colors to them so they can be differenciated easily on the dashboard. 

And there it went, created everything it needed, gave me the commands to build and test it, and it was working! I couldn't believe that I copy-pasted in Google Sheets for YEARS before I thought about this. Of course it's not perfect, sometimes I faced bugs while building or the app crashed in certain interactions, but I simply pasted the error and the context back into it and it would usually fix the problem in a matter of minutes. There was only one step missing from the dream I had where transactions would be auto-classified upon entry.

Making our app interact locally with the LLM

The last step was to connect the app to the local LLM installation (in my case Claude) so that it could classify the transactions properly. I also wanted it to cache the results when I edited the classifications so it wouldn't make the same mistake twice and it wouldn't use those precious tokens for no reason!

A couple prompts later, it was working, gave me instructions on how to make sure Claude is installed in the CLI beforehand and sent me on my way. The rest is history.

From then on, I continued to iterate on the app, adding multi-budget planning capabilities, user-management, authentication, password reset, improved the dashboard, and many more features that are still not pushed to the official branch. This was the beginning of a love story that would soon reach new heights. And with that, Family Finance was born.