Natalie Geller is the founder of Leevli, a real estate technology platform connecting home seekers with resident insights. Trained in law in Brazil and the United States, she has worked on major legal matters involving technology, including a U.S. Department of Justice case involving the Foreign Corrupt Practices Act (FCPA) and one of the largest settlements of its kind.
Natalie, you had a real estate broker license in New York and you are also an attorney in New York and Brazil. What gap in the home-buying process led you to build Leevli?
Finding a property has become much easier. People can search listings, compare prices and see a growing amount of data about a unit. Understanding what it may be like to live in that building is still hard. How does it operate day to day? How does management respond when something breaks? What do the documents say about rules or planned work?
As a broker, I could answer every question on the listing and almost none of those. My legal work taught me the other half of the problem: no single document tells the whole story. A listing, a disclosure, a public record and a resident's experience each reveal something different.
Leevli grew from the idea that buyers and renters should be able to see more of that context before they commit. Most AI in real estate has been applied to pricing and lead generation. We wanted to apply it to the question the resident is actually asking: should I live here?
Leevli describes itself as "the resident intelligence layer for U.S. residential real estate." What does that layer sit on, and why build a layer rather than another listings portal?
The foundation is nationally licensed MLS listings. On top of that we add verified resident reviews, the Leevli Score and Ask a Resident, a question and answer feature grounded in what residents have shared.
A new listings portal would compete on inventory, and inventory is largely a commodity. What is scarce is context: how a building is run and what the people inside it experience. Building a layer let us focus on that context while every listing still carries the facts buyers expect.
Today Leevli covers 12+ states, including Florida, Texas, Illinois, New York, Massachusetts and the District of Columbia, with more than 1,500 deeply reviewed buildings. That coverage also matters for international buyers, who often purchase in the United States without being able to visit a building or talk to its residents first.
Before launch, your team processed roughly 58,000 reviews across more than 1,600 buildings in over 10 markets. What did that work teach you about which signals matter?
It taught us to start from the decision, not from the data. Someone researching a building before buying a condo wants to understand management, maintenance, noise and how the building is run. Someone researching a neighborhood before moving cares about commute, daily routines and what the area feels like after dark. A review is useful when it speaks to one of those decisions.
It also taught us that volume is not the same as signal. "Great building" tells a buyer very little. A specific account of how a building handled a problem tells them a great deal, because it describes behavior, not just sentiment.
The most important lesson was about framing. A resident's comment is a perspective, not a universal fact about a property. The goal is to make firsthand experience easier to find while letting buyers weigh it alongside every other piece of evidence.
Resident reviews can easily become unreliable. How do you think about trust in resident knowledge?
The connection between a person and the place they describe matters. Knowing a neighborhood as a visitor is different from living there every day, and readers should be able to understand that context. That is why Leevli is built around verified resident reviews rather than open commentary.
Verification has a cost, and the cost is volume. Stricter standards mean fewer reviews, and some buildings will show thin coverage for longer. I think that is the right tradeoff. In a decision worth hundreds of thousands of dollars, honest thin coverage is more useful than a page padded with content nobody can stand behind.
No single review should be treated as the final word on whether a home is right for someone. Its value comes from being one clearly labeled perspective among several.
The Leevli Score rates buildings across eight dimensions using external data sources. How do you keep a score from becoming a verdict?
By treating it as a summary of evidence, not a conclusion. Eight dimensions is where we found the balance between nuance and readability. A building can be strong on amenities and weak on management, and a single number would hide that. Too many dimensions, and nobody reads them.
We use external APIs for the underlying data because those providers specialize in collecting and maintaining it. Our contribution is combining that data with resident experience, not rebuilding datasets that already exist.
The principle I care about is that a score should open a question, not close one. If a dimension looks weak, the buyer should be able to see what it is based on and decide what to ask next. A score that cannot be explained should not be trusted, including ours.
Ask a Resident answers questions using a retrieval-augmented architecture. Why retrieval, and what should the system do when it does not have the answer?
Ask a Resident runs on a semantic retrieval-augmented generation layer on Microsoft Azure. We chose retrieval because the knowledge changes constantly. New reviews keep arriving, and a model that answers from the current material for a specific building stays closer to reality than one trained on a snapshot.
The harder design question is what happens at the edges. A useful answer needs more than fluent text. It should point back to where it came from and make uncertainty visible.
There is also an important difference between "we did not find this in the available reviews" and "this does not exist." A system that blurs those two statements will eventually mislead someone. In real estate, a plausible but wrong answer about pets, parking or noise is worse than no answer at all.
You are building a legal intelligence capability for property analysis. What is the right role for AI in property due diligence?
AI is well suited to the repetitive, information-heavy part of the work: locating relevant passages, organizing records, comparing dates and terms, and surfacing details that deserve a closer look. That can make a complicated property file far easier to navigate.
For the first version, we made three deliberate choices. We work from public records only. Structured data is the backbone for factual flags, and we read document images only where they matter most, the deed and recorded declarations or covenants. The output summarizes, flags items worth attention, suggests questions to ask, and supports follow-up questions.
My legal training sets the boundary. This is information, not legal advice. The goal is for a buyer to walk into a conversation with their own attorney or agent better prepared, not to replace that conversation with an automatic verdict.
Why does property identity and scope matter so much in AI property due diligence?
Because a correct fact about the wrong property can still mislead someone. A record might concern an entire building while the buyer is asking about one unit. Another document may cover a different time period or a related association.
Before drawing any conclusion, a system has to keep those distinctions intact: the address, the unit, the building, the source and the date. It is one of the least visible parts of applying AI to real estate, and one of the most important. A fluent summary that merges two properties is worse than no summary.
What should happen when records and resident experience disagree?
The disagreement should stay visible. If two sources tell different stories, the buyer should see what each one says and where it came from. A system should not silently choose a winner.
Records and residents also answer different questions. A document can describe a building's rules and history. A resident can describe what living under those rules feels like. Neither substitutes for the other. When they conflict, that conflict is often the most useful thing a buyer can learn, because it tells them exactly what to ask before they commit.
What advice would you give to executives deploying AI in traditional industries like real estate?
Start with the decision people are trying to make and the information they actually use. Then separate the work AI can do reliably from the judgment that needs context, expertise or lived experience.
Ask where the model's knowledge comes from, and whether you would be comfortable showing that source to your customer. Much of the AI in real estate today generates confident text from listing descriptions, which were written to sell. If the inputs are marketing, the outputs are marketing with better grammar.
Finally, treat trust as the product. Features can be copied. Being clear about sources and limits is much harder to replicate. In a decision as consequential as buying a home, "I do not know yet" can be more useful than a confident answer built on incomplete evidence.
