romanleeb

smart

Hey HN,

I built Billpal because every tool I tried wanted to turn me into an accountant.

I'm self-employed, I don't need much – no P&L, no balance sheet. Happy to pay a CPA for this and to file my taxes.

But I also didn't want to do bookkeeping myself either. It's a stupid task.

Was genuinely surprised how little automation these accounting tools have. Even when I upload a receipt, I still have to type most of the fields by hand.

So I built my own thing: grab invoices from my inbox, parse them, and match each one to the right bank transaction automatically.

Now all I do is export a CSV and send it to my tax advisor.

Curious what you think. And happy to get into the weeds on the matching logic.

Roman

amazing

Thank you! Well, I built it in a way that the format is kinda universal – expenses are categorized but not with tax codes. Instead I am using labels like e.g. 'Cloud Infra & Hosting'. Plus apart from the basic infos I even extract things like reverse charge y/n or validate VAT IDs to make it as easy as possible for my tax advisor.

So it will not file your taxes for you but it prepares the raw data for your tax guy in a way that they have minimal effort and you can save quite substantially on the bookkeeping part. I'm in Austria and it is a bit cheaper here but also quite costly. Now I only pay for the tax filing, annual accounts and consulting when I need it.

Hey HN,

I'm Roman and I've been working on Billpal, a tool I built to do my bookkeeping for me.

Started as a personal thing, but I ended up building a product around it so others can use it too.

I'm a freelancer myself and always had a strong dislike for accounting, so I started to automate things bit by bit.

After testing a lot of accounting tools, I realized that a) they have many features I don’t need, and b) I was still faster using a spreadsheet.

So I built this tool to focus only on the essentials: finding invoices in my inbox and matching them to bank transactions.

What it does:

- Imports bank transactions via Open Banking APIs

- Collect invoices from the user’s inbox

- Using an IDP/OCR provider for document extraction

- Extracts basic info like date, supplier name, and line items, as well as specifics like whether it's reverse charge or various tax rates (e.g. from restaurants)

- Matches invoices to expenses based on params like amount, currency (FX conversion), and similarity to supplier names / line items (that's the only agentic part)

- Can also handle credit card statements (this part still needs improvement though)

- Users can then export a CSV with matched records

- Also built a whatsapp bot to make it super simple for users to upload receipts (photo or pdf)

Currently supports banks in the EU and UK, but planning to add US banks down the line.

Focusing on expenses right now but will also add income shortly (outgoing invoices from sent emails).

I would love to get your feedback on it. Check it out here: https://billpal.io

Looking forward to your thoughts / suggestions,

Roman

Thanks - we'll do!

I agree with you mostly but not 100%. It's definitely true that the way how you build it makes a product stand out - however the goal of this tool is more to understand WHAT to build as it gives you a glimpse into the opinions customers have about other (similar) products. The how is then onto the devs for sure. Yes you are right if you know for instance Hubspot you do not need this application to tell you stuff about Hubspot you already know, it tells you how the audience perceives the product and the offering.

Data sources are various sites that have detailed reviews on software. We did not fine-tune anything yet as the initial goal is and was to validate the idea per se. We'll improve it further of course once we see there is a need for this.

Very good input, thanks! We'll add some examples on the page.

Currently we have a quite basic RAG setup, no fine-tuning.

We wanted to get this out fast so decided to use Assistants API but definitely needs some polishing in this RAG pipeline (e.g. having more control over the exact search queries). For the MVP it was convenient as it all comes out of the box. Metadata filtering also better with Pinecone for instance.

Precisely, problems/complaints/negative comments unveil a lot of interesting opportunities.

For this very reason we instructed the model to pay a bit more attention to expressions with a negative sentiment.

That was exactly the reason we built this - to facilitate the gathering of this intel. Will it be a panacea for doing research - of course not. But it can speed things up a bit and throws another (probably faster) research methodology into the mix.

If you're interested we wrote a longer article about how this idea came to be: https://medium.com/@roman_leeb/how-to-figure-out-what-users-...

I agree, there is a lot of opportunities you can find when looking at a subset of frustrated customers. Ironically they tell you in greater detail what they dislike vs the super happy customers on why they like something.

Hm can you elaborate on this - I do not fully understand. Was there no response for "Summarize the reviews you have on products in the OCR category"?

The demo is not "cached" or anything this is production.

Would be cool if you could give us more info on this.

So the difference is we inject software review text chunks into the conversation as hidden context for the LLM to answer the query. Based on your input we do a cosine similarity search in the vector DB and retrieve the most relevant results which are then analyzed and subsequently the model formulates a response.

Maybe we need to make this even more clear on the LP - like the comment earlier where we show responses side-by-side from e.g. plain vanilla OpenAI/Claude and Reviewradar.

Currently we do not have a function caller that checks the topic of the prompt and then direct the query accordingly - this is a very good input thank you! We did some work to add relevant details to the review chunks like e.g. the category of the tool in question but we know we got work to do here to make this even smoother. Kudos for bringing this up here, we'll 100% look into what you suggested.

This is a very good idea, thanks a lot for this!

Yes we do intend to monetize this at some point, currently offering access as a free beta to first validate whether this is useful as well as learn more about the specific prompts that the model is expected to answer (and then obviously improve it based on that knowledge). But what you proposed with side-by-side comparisons on the LP itself is a fantastic idea.

Hi this is Roman, the co-founder Andrei mentioned in the post.

I agree with you, a good chunk of these reviews come about through incentivizing people and not all of them are organic in nature. However, many are genuine and even the ones that are from people receiving some sort of credits do contain some very interesting things. But I get your point, would be better if there was no carrot in front of their faces urging them to leave a positive review.

Also we do not take them at face value but instruct the model to distill out the essence of what is mentioned when analyzing the user feedback and formulate a response based on patterns embedded in the findings across multiple reviews.

Curious on what you got as a response from Claude - can you share the prompts with me?