Absolute n00b trying to get started building an AI agent to do a weekly budget report

General
1

Hi Guys,

Nice to meet everyone. I am very new in this space, as a pilot project to get understand workflow and targeting a use case that makes life easier.

As a dad, having an agent that logs into various banking. billing accounts and download a monthly csv. formats data and aggregates based on some other parameters and outputs a report like how our family usually reconciles would be an amazing start.

Do you know how abouts I would start this journey?

Thanks,
Kabi

2 Likes
2

as a fellow noob i would suggest asking free chat gpt 4-o that exact question, ask it to create a series of prompts for an AI coding agent to use to build it, then copy each prompt and feed it back into gpt, and say can you please think carefully about this prompt and anything we may have missed and expand on the prompt?

then use those prompts in bolt with the model of your choice

2 Likes
3

Wow, noob party. I’m +1 in the same category. I am a spreadsheet guru (self proclaimed of course) but would really like to automate it also. Would be nice to be able to send csv-s through a system that can check for overlap with previous, then exclude all the overlap and replace the pending transactions appropriately. Too bad mint.com got taken down, that compiling transactions and allowing you to download was my bread and butter till they disappeared :sweat_smile:

4

Agree to @jing and I would split it in backend / frontend.

The backend part you could do with python as it is fast and runs on the server side.

  • You could utilize “Cursor” for Implementing it. Provide it some example data, let it write into a database (could be just a sqlite or even a jsonfile for beginning) and provide API Endpoints

Frontend you can build with bolt and tell it your requirements, as well as the response strcuture from the backend.

2 Likes
5

also forgot claude sonnet 3.5 has a free browser chat plan , milk both those free daily allowances to get the prompt designing done then use a cheap api to do the legwork imo

1 Like
6

break it down into parts, as a noob it might be difficult to implement a system that logs into banking accounts considering the level of security against bots that banking accounts have. how i would go about this for faster implementation would be:

  1. download the files yourself and save to a folder
  2. have the agent format teh data and aggregate to give raw analysis (pure python df manipulation using numpy, pandas, some visualization using matplotlib or seaborn)
  3. have the agent pass the raw analysis (all dfs, optionally the visualizations as well) to a llm to analyse, give business insights, reviews, points to consider, outliers etc. and format it as a monthly report. i would recommend google gemini flash 2.0 for this due to super large context window being abe to fit more of the dfs.

You will have to work on step 2 a little bit to ensure all important data is being captured. it is difficult to pass entire csv if the file is too large to an llm as when its embedded it will exceed context length. I built the same workflow at my workplace and ran into the same problems.

Hope this helps!!! please let me know if you have any doubts, or pass this entire message to chatgpt and have it create a plan

7

Thank you all for this! I will try these suggestions and provide an update of how my venture goes!