Track personal finances nightly from bank emails

Tools

Claude Code

Overview

This is the closest I have come to "feeling the AGI"

Workflow

This is the closest I have come to "feeling the AGI" I suspect there would be a sharp division in the debate on AI between those who have used Claude Code and those who have not. I use the term loosely here, as a placeholder for any frontier agentic coding tool. For a non-techie, I have been following the post-LLM era of AI fairly closely and have always been more bullish on AI than the average person. Overall, however, I have been somewhat tempered compared to fans, which is probably in the same space as Arvind Narayanan-Sayash Kapoor level of excitement. Which is to say that the current progress is very impressive; I see lots of potential, but I am not completely awed by it as if it were the ‘god machine.’ At a conceptual level, I am still broadly in the same place, but my sense of awe has considerably increased after playing around with Claude Code over the last two weeks. It is the most AGI-like thing I have ever experienced. So far, I have been occasionally impressed with chatbots. I have tried various prompting techniques and most fads. But the results often lacked something. Mostly, it did not have a great sense of which details are important and reliable. And the writing is always too… AI-like. For better or worse, I have developed a very strong sense of AI-generated prose, and I am irrationally repulsed by it. But I feel differently after using Claude Code. Most of the time, it can actually pick up on the essence of what I want; it goes on for hours sometimes; it tries, fails, and troubleshoots problems on its own. I do not know much about coding beyond some data analysis for policy research. Yet, I have often been able to get what I want. Here are some random projects I have worked on in the last two weeks or so. 1. Dino internet race This was completely Simran Saini’s idea, which I worked with Claude Code to execute. During video calls, whenever there is lag, I jokingly tell my friend(s) on the side to fix their internet. And they would retort with something along the lines of “actually, it’s your network that sucks.” This is all just a nonsensical talk-and-forget type of conversation. Dino Race allows you to settle that dumb debate by ‘racing’ against your friends (up to six people) live, where your dinosaur avatar runs as fast as the speed and stability of your internet. The person with the best internet wins. It is purely for fun, and you can try playing it here. Also, here is the GitHub repo. I am not aware of anything similar to this. Claude Code completely built it from scratch. The URL is a little strange because it is a free subdomain. My personal Raspberry Pi acts as the game server. Having a home server gives me a strange kind of cheap thrill. But why do I own a RasPi in the first place? This brings me to the next project. 2. WhatsApp message scheduler Sometimes I want to reply to people, but the timing feels odd. These are usually semi-professional messages, and the recipient is in a different time zone, or I checked the messages too late at night. Often, I just want to schedule messages in advance. There is no way to schedule messages on WhatsApp natively, and I hate using third-party integrations with my personal communication apps. So, Claude Code helped me build a scheduler. I had to log in to my WhatsApp Web once, and a script operates a headless browser. The obvious problem was that my PC needed to be active when a scheduled message was supposed to be sent. That was impractical, so I bought a Raspberry Pi 5. It is a low-power single-board PC that I had been considering buying for some time, but had no obvious use case until then. I later added a bot option that lets me message my Telegram bot to schedule a WhatsApp message.1 I can use natural language or a fixed format, and it will update my WhatsApp Scheduler. The bot processes my natural-language instructions using a local LLM (that runs offline on my Raspberry Pi)2, so I do not even spend any tokens for this. The WhatsApp-Scheduler GitHub repo is available here. 3. WhatsApp CRM Since I have my WhatsApp web logged in on the headless browser on RasPi, I thought it might be interesting to do a regular ‘content analysis.’ The WhatsApp-CRM scans my messages and, based on context, classifies contacts as personal, professional, or service. Excluding one-off transactional or service-based conversations, it surfaces: Dormant chats: a nudge to connect with people I have not spoken to in a while.

Reported result

The tracker successfully aggregates expenses across accounts and adds merchant context automatically. The author notes it is doing an incredible job, though they admit to rarely checking the reports due to a preference for avoiding spending awareness. No quantitative metrics on accuracy or time saved are provided.

Human role

The original post does not say how a person stayed involved.

Limitations

Requires a Raspberry Pi running a local LLM (Llama 3.2:3B) and access to bank email feeds. The author warns that as a non-coder, they cannot judge code elegance or scalability. Integration with specific bank email formats may require custom parsing logic.

Try it

Act as a technical assistant to build a personal finance tracker. 1. Analyze sample bank notification emails from [replaceable input: email samples] to identify transaction patterns. 2. Write a Python script using a local LLM to extract date, amount, and merchant, then categorize spending. 3. Set up a cron job to run this script nightly. 4. Output a JSON summary of daily expenses. Before finalizing, run the script on [replaceable input: 5 test emails] and show me the extracted data for approval. Do not deploy until I confirm the categorization logic is correct.

More cases · Finance & legal

• Investigate market moves and alerts with an agent on your charts
• Fight consumer claims and recover refunds
• Altumatim: Electronic discovery
• Cancel and refund subscriptions by searching email and calling
• Build a spending-insights app from bank transactions