I've lived in the UAE for 3 years now, and it felt inevitable to consider investing in the real estate space here, partly to own a home that I could live in, but also because the market here is truly rewarding. So when I started my search, my primary objective was to find the "best" housing opportunity. And soon, "best" became quite difficult to define. Is it somewhere I see myself living in - green community space, access, decent size and layout? Or is it good value - price per sq. ft., rental potential, payment plan? Every deal looked different on paper, and my reasons to buy kept changing with it. Was that the right approach? Absolutely not!
Every deal that my broker(s) sent me went into Claude, and I asked various questions: 1. What's the IRR? 2. What's the rental yield? 3. How does this compare to other properties in the area? 4. Is the layout Vastu compliant? And so many more. And every time, Claude gave me a reason to reject the deal. I took a pause and truly asked myself why I wanted to buy a house, and forced myself to prioritise. In a classic conflict of balancing finances while achieving my objective, I realised that it's OK to have 2 goals for the same task. My method of evaluation, and my parameters, would change. By this time, I had gathered a lot of data, questions, and had trained Claude enough to create an agent that could assist me in evaluating each deal and making the complex decision.
Here's What I did
I handed it to a model and worked alongside it, feeding floor plans, sales offers and screenshots of a 3D sales viewer, one at a time. Every time I assessed the output the model gave me, I also understood every place the process nearly went wrong. This was the real output, a blue print to my agent. To build an agent that is precise, I do not have to imagine what it should do, I simply encode the corrections I made by hand. Mistakes mostly lied in the source documents, leading to misinterpretations.
For instance, the north arrow on a floor plan is the arrow for the drawing sheet, not for the building on the ground. There were times I had to tell the model how to read an illustration, and use the compass. Area meant three different things at once, and a flat that looked cheap turned out to be forty six percent terrace. The headline discount is was the entry price, not money already made, so counting it as future growth was counting the same gain twice. On occasions the model hallucinated and invented its own winning deal (an apartment choice which didn't exist).
I had to cut out the prose and make a clean input/output plan. Two hours later I had a decision I trusted, based on parameters that I chose, and priorities that truly mattered to me.
What I built
Three parts - 1. A form that says what to collect about a flat. 2. A small calculator that turns those facts into yield, cash and a Vastu score. 3. A reading assistant that pulls a messy PDF into the form.
The idea is that the model reads documents and writes the argument. A plain engine runs the maths and produces every number, so the output is the same every time and I can read the four lines that produced it. A human confirms the one input that is convincing and often false, which is the compass. Nothing is inferred to fill a gap either a field is traceable to a source or it is left empty (no room for an invented answer).
Try My Agent
The demo takes a sales offer, a few numbers and the compass, and scores a property on the three things that actually decide it: whether it is good to live in, good value, and a good return. Upload an offer and it reads the price and the area for you. When you move the sliders, the dials move. Every figure shows its working.
It is a heuristic, not a verdict and the weights are my starting judgement and the tool shows them so you can argue with them. The assumptions sit on the page, and the compass (the real pain point) is yours to confirm. None of the output is advice, it clearly shows how the deal ranks against the multiple goals I want to achieve, the deciding factors to help me decide the sweet investment spot.
Where this goes
This is basically a version zero - it does the job it was built to do. The next step is wiring in live rent and price comparisons so the benchmarks don't depend on manual searching, plus a watcher that can pre-score new listings against an ongoing brief.
But the important thing is that the shape of the system came from actually doing the search, not from sitting down and drawing a diagram first. I think that's the bigger lesson: you can't really design an agent for a job you haven't done carefully yourself. Once you've done it, the design is mostly a record of the mistakes you made and the things you had to fix.
