Hypocalc - Developing mortgage calculator with llm dev tools

Introducing Hypocalc

Hypocalc began as a toy project. I wanted to see if I could build something useful entirely from scratch using LLMs. No templates, no frameworks I was already familiar with - just me describing what I wanted and AI helping me code it up from zero.

The result is a financial calculator that compares two common investment paths: buying property with a mortgage versus renting and investing in the stock market. It goes beyond basic mortgage calculators by looking at property appreciation, interest rates, down payments, rent increases, and stock market returns.

People compare renting versus buying all the time, but they rarely factor in stock market returns. I wanted to test this with different variables - what happens if the property market grows slower than expected? What if stocks perform better? What if interest rates change? Building this tool from nothing using AI was my way of answering these questions with real numbers.

Try it yourself at hypocalc.app

The Development Journey

Hypocalc served as a test bed for LLM development where I could experiment with AI coding tools for frontend work - not my usual territory as a backend developer. I basically wanted to experience coding with LLMs as someone who doesn’t know how to code - because I pretty much don’t know frontend.

LLM-Assisted Development

I tried several different approaches with AI tools:

Copy-Pasting from LLMs

Small steps - the workflow everyone was doing three weeks ago. I’d describe what I needed, copy-paste the code, run it, check the results, then go back with questions. The loop was too long, especially since LLMs wouldn’t repeat all the code in their responses. It worked, but created a messy patchwork that was hard to maintain.

Avante.nvim

Avante.nvim is a plugin inspired by Cursor, and interestingly I tried it before checking out Cursor itself. It’s good - it has APIs to many models. It was very useful with DeepSeek R1, before Claude 3.7 was a thing. It doesn’t always know how to apply the diff well though. I read that Cursor has some proprietary model for the diffing in the editor, and it shows. Avante works a lot of the time, but it’s not nearly enough to make you forget the tool and focus on development.

Claude Code

This is the king - I struggled with calculations for a long time, and Claude solved them in one shot. It costs some money and can burn through $10 in a prompting session. Super impressive though. It also asks for permissions way too much. I’d like to dockerize the environment for one of my projects so that it doesn’t need to ask for confirmation on every small thing.

Cursor

Really good, pretty much like Claude Code but with more options to intervene. Its main advantage over Claude Code is the price. You pay $20 and get unlimited requests, even though they can be slow after some time. I suspect they have some deal with Anthropic and OpenAI that makes their API price way cheaper, allowing them to just charge $20 or $40 for a subscription. I don’t like the VSCode browser, but you can use Cursor just for prompting and switch to your preferred text editor when needed.

Adversarial Development

I stumbled on a neat trick I started calling “adversarial development” - using different AI models to check each other’s work.

Here’s what I did:

1. Took screenshots of Hypocalc’s interface
2. Asked ChatGPT (the vision version) to critique the design
3. Fed that criticism to my coding assistant to fix the issues

This worked surprisingly well:

• The vision model caught UI problems that text-only models missed
• Since it didn’t write the original code, it had no problem pointing out flaws
• Each round of feedback made the app noticeably better

It reminded me of how GANs work, with different neural networks playing off each other. I was basically using one AI to build and another to critique.

I think we’ll see more of this kind of AI adversarial setup between different models. It really helped me, as a backend dev, make something decent-looking on the frontend.

Fetching Data

For a project like this, it’s important to know the data - average rents in Prague, typical apartment prices, and current mortgage rates. It’s awesome that LLM deep research features can help you with just that. The calculator is based on real data, not just guesswork. I can even set up a cron job that will update interest rates automatically (haven’t done that yet, but it’s on the roadmap).

Try It Yourself

Hypocalc is free to use online. All calculations happen in your browser - no data gets sent anywhere.

If you’re wondering about whether to buy or rent, give it a try and see how the numbers look over 5 to 30 years.

Check it out at hypocalc.app

Note: While Hypocalc provides powerful analytical tools, remember that any financial decision should consider your personal circumstances, goals, and risk tolerance. The calculator offers insights based on historical data and mathematical projections but cannot account for all variables in financial markets or predict future events with certainty.

Key Takeaways:

• LLM tools change so quickly, even during one project I switched many times
• Combining multiple AI models creates better results than using one
• Frontend development is accessible to backend developers with AI assistance