Taxpal - Localized Tax Prep Assistant

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

I have built a tax-prep assistant (I have decided to call it taxpal). It basically takes in your bank statements -> sorts, checks them and
ultimately does the math producing the worksheet for your accountant to file.

Now why would anyone pay for it, well for starters if anything is wrong in your Schedule C, IRS could send you a notice asking for an explanation or pay more tax.

Sure, isn't that what I pay my accountant for ? True but accountants don't do that for free do they, they could charge you around $300–$1,500 for sorting a year of transactions, something taxpal could do for $10.

Demo

This is what the website looks like!!!

Code

% https://github.com/blixygetir/taxpal

How I Built It

The service uses Hermes 3 (Llama 3.2, 3B) which runs in the browser with WebLLM via WebGPU. The model does make mistakes due to its smaller size and the design compensates for it.

A harness drives the model, it ensures that the reply must be JSON schema. A JSON schema per tool restricts the model to its real tool names.

A RAG looks up into the actual IRS documents and makes sure that since the model does not reliably know the IRS rules, it won't fake

the citations.

A loop sends the past 6 of the user's reply to make sure that the model learns from it without getting trained over and over again.

Why Does Open Innovation Matter?

The app uses a local model which runs on the browser (via WebGPU). The data never leaves the user's machine, so the LLM is not trained on user's data as well. Primarily it protects the user's Financial Documents as if its life depends on it :) !!!