blog.rudel.nl

^{or: how I stopped fearing AI}

finding a new home is hard.. not only do searching WHERE you want to live.. and IF you can find work there.. but also

will I have space for my dogs
how many neighbours will I have
is it noisy? train lines? aircraft noise?
can I immediately live there or do I have to renovate?
can I walk my dogs? busy roads nearby? will there be water?
and so on… a LOT of factors that define happiness. and only so many can be asked during a visit. Also.. the internet has a LOT of websites trying to sell you property.. but which ones are worth investigating?

enter GPT-ImmoRadar

it started as an idea.. what if AI could understand what I want and help me create a shortlist of objects, that I want to investigate.

the basis is an email search agent from the mortgage agency / broker. it is then enriched by geodata, water, neighbourhood, water, noise, etc.. data that I can publicly find. furthermore a browser plugin pulls data from the expose, like price, condition, energy label, etc.

this is not AI doing my job, this is interaction between GPT and a human, massively improving my capabilities, saving me a lot of time and.. honestly.. I haven’t had so much fun with a self-hosted project in years. oh yea.. it is self-hosted, I don’t think anyone would like to do this and not make money with it. I don’t care. download the code.. adjust the scoring algorithm for your personal situation, find the house of your dreams.

I was afraid if AI. Now, I am still (or even more) afraid.. but I have seen what a powerful engineering tool it can be.

Not by asking ChatGPT “build me an app” once (I admit, thats how it started) but through hundreds of iterations: requirements, implementation, production tests, failures, debugging and architectural changes we have now arrived here:

AI did not remove the engineering work. It dramatically increased what one person could build.

interaction with the server via telegram and browser

in the end I can create a top-3 list that I can investigate deeper.. ask for the exact address, rescore land, water and noise levels.. contact the broker and plan a visit.

the “workflow” is like this:

Property alerts
|
Email parser
|
SQLite -- Geo / Water / Noise
|
Telegram
|
Human + Browser + Tampermonkey
|
Browser enrichment
|
Explainable scoring

the “scoring” is based on my preferences:

For example, a good house might score like this:

Price + condition       19 / 25
Plot                    16 / 20
Neighbours              17 / 25
Quiet                   15 / 15
Water                    8 / 10
Living area              4 / 5
Energy label "A"         +2      <- bonus
--------------------------------
Total                    81 / 100  -> very interesting

Energy is deliberately shown as an adjustment rather than as a fixed category worth a certain number of points. A good label can add points; a poor label can subtract them.

A different house can look cheap at first but still rank much lower once condition and energy are taken into account:

Purchase price          130,000 EUR
Condition surcharge     +140,000 EUR  (major renovation)
Effective price         270,000 EUR

Price + condition        9 / 25
Plot                    11 / 20
Neighbours               6 / 25
Quiet                   12 / 15
Water                    7 / 10
Living area              3 / 5
Energy label "H"         -8      <- penalty
--------------------------------
Total                    40 / 100  -> archive

In the current model, for example, energy label A gives +2 points, while label H deducts 8 points. The detailed energy adjustments are configurable and documented with the rest of the scoring model.

source code is public

I don’t like proprietary stuff.. also I thing GPT and immowelt/immobilienscout earn enough. this is just me using what is out there so.. use it. find your dream.

https://github.com/flamingsnowflake/GPT-ImmoRadar