Hovis - the real estate expert's copilot
For REI School in the US, I built Hovis, an AI copilot trained on one real estate expert. Students get his answer at any hour across 14 courses and 328 lessons. Every deal number is computed in code, 213 checks and none failing. Built on Claude, Supabase, Railway, GoHighLevel and ElevenLabs.
The Challenge
One expert cannot answer every student, and the questions arrive at 2am. The risk was the fix itself: an AI that sounds confident while being slightly wrong is worse than no AI, because a student can act on it and buy a house. An early version did exactly that. A missing operating cost made a losing rental look like $452 a month of cash flow, and it told the expert to buy it.
Results
- The expert’s own test deal had shown $452 a month of cash flow that was not there. It now returns a walk.
- Students get a real answer at 2am instead of waiting for the next group call.
- One expert now covers 14 courses and 328 lessons at once, with no second hire.
- Students see the honest net return, not a gross number labelled as net.
- Deal numbers are computed in code, not written by the model: 213 checks, none failing.
The Solution
- 1
Every deal number is calculated in plain code, not by the language model. The AI narrates the result, it does not compute it.
- 2
A fact only ships if the quote behind it can be matched back to the expert’s own material, so the AI is not free to put words in his mouth.
- 3
A missing deal input becomes a question back to the student, not a favourable guess.
- 4
Answers cite the exact lesson timestamp or book page, and only unlock what that student actually paid for.
- 5
Spoken answers run through number-formatting checks before release, so figures are not mangled out loud.
Want a system like this in your business?
Book a free call and I will show you exactly how this would work on your stack.
