Date: 06 Oct 2024
Ops suck! Recently, I have started to keep a record of stuff. One such requirement was keeping records of all the spending and maintaining a Google sheet of the expenses. I decided to keep the invoices in a Google Drive folder and maintain the sheet with the GDrive link to these expenses. Though, it is what I needed; there was one little buggy caveat to this process. The pain of looking up the data in the invoices and entering them manually in the sheet.
There are solutions like Zoho, but it seemed like an overkill for my requirement. I just needed a solution that could take out a few required fields from invoices and put them on my sheet. Zapier here seemed like a good solution, but it is quite limited for the niche solution I was looking for. Having LLMs available today is like having a personal junior software developer, using which I was able to complete the entire pipeline in about 1hr.
Leveraging existing automation solutions
Zapier has a very neat way of creating automation pipelines. Since the task of uploading invoices to a Google Drive was straightforward, I created a zap to automate this. This was configured so that when I send an invoice on a Slack channel, it would get uploaded to the designated Google Drive folder. This solved one part of the problem for me. I now needed to increase the capability of this automation to extract the required fields from different types of invoices and push them to a Google Sheet.
Enhancing the capabilities with LLMs
I gave ChatGPT a prompt
“””I have some invoices in a Google Drive folder. The required data is filled out in a Google sheet. These are the columns in the Google sheet "Date, Item Cost, Invoice link, Refund"
Date is the date of the invoice
Item Cost is the cost
Invoice link is the Google Drive URL
Refund is a boolean field indicating if it is a refund
If the refund is true, then the cost is updated as negative. I need you to write a python script that automates this. The invoices could be both pdf and png.”””
This gave me a code that was doing 60% of the required job. A bit of minor tweaking and I had set up a pipeline to automate 80% of the task. For a few invoices, it was unable to get the required fields but it was good enough for me.
The complete pipeline
For automating invoice upload, Zapier was great to set it up for my use case where if I upload a file to a particular Slack channel it gets uploaded to GDrive. Which took me about 10 minutes. Then to extract the information from the uploaded invoices, ChatGPT wrote me a Python script, and with minor tweaks, it was a good enough automation for me. Overall, it took me just about an hour to set it up. You can reuse my implementation for data extraction from here: https://github.com/Telwha/invoice-data-extractor
My comments
This new era of AI and LLMs advancing day-to-day has made it incredibly easy to achieve our goals rather than fiddling around to find solutions. It has made it so easy to build automation for our day-to-day tasks and focus on crushing goals. Even if the solution is not well-rounded, it still saves a lot of time, and IMO, a reduction of 80% for my day-to-day ops tasks is huge for me!