Real Estate Intelligence

Turning real estate data into an intelligent acquisition and decision system.

Illustrative exampleNot published

01 · Business problem

A national real estate network has fragmented inventory, supply spread across the territory and buyers who say what they want in sentences, not in filters. This case shows how Menlo Tech turns that problem into a system: it interprets intent, crosses it with a geographic graph and with properties, ranks options with explicit criteria and coordinates agents that publish content and learn from the market.

The method, flow and agents are presented in a demonstration environment and in the tour (illustrative example). Measured client results are not published on this page.

RE/MAX operates as a network: offices and brokers across the country, each with its own portfolio, and a brand that must be found by people looking for a place to live.

For the buyer, the right property depends on place, price, commute, school, neighborhood and routine. For the network, it depends on turning that local knowledge into useful pages and answers, at scale.

The operational problem

Search by preset filters (neighborhood, price, bedrooms) misses what people actually say: near the metro, no line changes, school within walking distance, and quiet.

Data is disconnected: properties, geography, search demand and content live in different systems, and every gap becomes manual work or a missed opportunity.

Existing process

Search by fixed filters, without interpreting intent.
Local content produced page by page, with manual review.
Inventory, geography and demand in separate databases.

Limitations

Compound intents do not fit in filters.
Scaling local pages depends on editorial hours.
Without a learning loop, what works is not repeated systematically.

02 · Mathematical foundation

Semantic retrieval

Illustrative example

The buyer’s sentence and each property become vectors (embeddings), and similarity measures how close they are. This is how “near the metro” finds properties described in other words.

Variables

ssimilarity between query and property (from −1 to 1)
qbuyer query
pcandidate property
Efunction that turns text into a vector

Objective

Generate candidates: keep the properties closest to the query so they can be ranked afterwards.

Constraints

Hard constraints filter first (for example, active listing and region).
Similarity generates candidates; it does not decide on its own.

Practical example

A query about an apartment near the metro to get to work brings up properties in neighborhoods with a station, even when the listing does not use those words.

Business implication

Fewer relevant properties stay invisible because of vocabulary.

Assumptions and limitations

Quality depends on the embedding model and on data coverage.

Multi-objective ranking

Illustrative example

Normalized criteria (location, price, transport, school, comfort, preferences) combine into one score per property.

Variables

Ssuitability of property p for query q
wᵢweight of criterion i (non-negative; sum to 1)
fᵢcriterion i normalized between 0 and 1
nnumber of criteria

Objective

Maximize perceived suitability while respecting hard constraints.

Constraints

Hard constraints remove candidates before the sum (for example, no line changes, when the buyer requires it).
Weights sum to 1 and may vary with intent.

Practical example

In the demonstration, changing the weight of walkable school reorders the properties and shows how much each criterion contributed to the score.

Business implication

The order stops being opaque: each position has a rationale and explicit trade-offs.

Assumptions and limitations

Weights must be calibrated with conversion evidence; the ones in the example are illustrative.

Constrained optimization

Illustrative example

Choosing what to publish, highlight or acquire with limited resources is an optimization problem: each action has an expected value and a cost.

Variables

xᵢ1 if action i is executed; 0 otherwise
vᵢexpected value of action i
aᵢeffort or cost of action i
bavailable effort budget

Objective

Maximize the total expected value of the chosen actions.

Constraints

The sum of efforts does not exceed the budget.
Only actions eligible under policy (quality and portal rules) enter.
Each action is either in or out.

Practical example

Among candidate pages, which to publish in this wave given the editorial review limit? The tour shows staged launch with stop criteria.

Business implication

Editorial effort goes where it can return the most.

Assumptions and limitations

The value vᵢ depends on demand forecasting; without measured data it is a hypothesis to validate.

Demand forecasting

Illustrative example

It combines recent history with the previous forecast to estimate demand in the next periods. Geography and seasonality enter as model extensions.

Variables

ŷforecast demand
yobserved demand
αweight of recent data (between 0 and 1)

Objective

Minimize forecast error.

Constraints

α stays between 0 and 1.
Seasonality and geography need their own models, beyond this smoother.

Practical example

If searches for living in a neighborhood rise every start of the year, the forecast anticipates pages and inventory for that period.

Business implication

It prioritizes launches where demand tends to grow.

Assumptions and limitations

Exponential smoothing is the simplest method, used here to explain the idea; the final choice requires validation with real data.

Journeys as a Markov chain

Illustrative example

Each visit moves through states (search, list, property, contact, conversion). Transition probabilities describe where navigation tends to go.

Variables

Xjourney state at step t
Pᵢⱼprobability of moving from state i to state j

Objective

Estimate the probability of conversion and locate where journeys are lost.

Constraints

Each row of the transition matrix sums to 1.
It assumes the next step depends only on the current state.

Practical example

If many journeys leave the list for outside the site, the priority improvement is in the list, not in the property.

Business implication

The learning loop targets the stage that loses the most.

Assumptions and limitations

It requires navigation data with consent; the example probabilities are not measured.

Run the calculation

03 · Technical architecture

Properties

Attributes that describe each property and its publication state.

Price, area and bedrooms
Location and features
State: active, in review, published

Geography

Territory entities linked in a graph.

Neighborhoods
Stations and lines
Schools and commute axes

Demand

What people search for and ask.

Search queries
Suggestions and questions asked to assistants
Grouped intents

Behavior

Signals that feed learning.

Navigation
Contacts
Conversion

Real sources and volumes depend on client-approved data. The tour and the demonstration use illustrative data.

Architecture layers

Data and knowledge
Property base, geographic graph and search demand.
Proposed architecture
Intelligence
Embeddings, intent classification and forecasting.
Proposed architecture
Decision
Multi-objective ranking and constrained optimization.
Proposed architecture
Orchestration and policy
Specialized agents and validation of each action.
Proposed architecture
Observability and learning
Event log, indexing, impressions and conversion.
Proposed architecture

Reference architecture: not every component exists in every deployment, and this case does not state which ones are in production.

04 · Agentic execution

Intent Agent
Turns the buyer’s sentence into structured intent, with hard constraints and preferences.
Property Matching Agent
Retrieves candidates by similarity and applies the constraints.
Geographic Intelligence Agent
Queries the graph for commutes, stations, schools and neighborhood.
Content Quality Agent
Checks that the page has the required blocks from data and that the text is useful.
SEO Validation Agent
Checks indexing rules, duplication and file limits.
Publication Agent
Publishes in waves, within the stop criteria.
Performance Analysis Agent
Tracks indexing and impressions and feeds learning.

Agents propose; a policy layer validates each action before execution. What an agent cannot do, such as publishing without passing quality control, is a rule, not a suggestion.

Execution workflow

Natural-language query
The buyer describes what they want in their own words.
Illustrative example
Intent classification
The sentence becomes intents: commute, school, quiet, natural light.
Illustrative example
Semantic representation
Query and properties become vectors for comparison.
Illustrative example
Knowledge graph
The territory enters as verifiable entities and relations.
Illustrative example
Property matching
Candidates are retrieved and filtered by hard constraints.
Illustrative example
Ranking and optimization
Criteria combine; trade-offs are explicit.
Illustrative example
Agent execution
Agents prepare pages and answers; policy validates.
Illustrative example
Conversion feedback
What converts goes back to ranking and prioritization.
Proposed architecture

05 · Business impact

Search relevance Not published
Compound intents are understood, not just filtered.
Lead qualification Not published
People arriving through a specific intent tend to arrive better prepared.
Inventory discovery Not published
Properties described in different ways become findable.
Conversion Not published
More fitting results should bring search closer to contact.
Content operations Not published
Local pages with blocks coming from data reduce repetitive manual work.
Acquisition efficiency Not published
Demand-based prioritization directs the network’s effort.

Each line describes a relationship expected by the design. None is a measured result: no percentages are published for this case.

06 · Evaluation & learning

Governance and human oversight

A policy layer validates each agent action before execution: page eligibility, required blocks coming from data, file limits and quality rules.

Human editors review what automation cannot resolve safely and decide exceptions.

Launch is in waves with explicit stop criteria: the next wave only opens if minimum indexing and the rate of pages without impressions stay within the defined limits.

Every decision leaves an event record that makes the path auditable.

Limitations and assumptions

Tour and demonstration data are synthetic or illustrative.
No measured client result is claimed.
Embeddings, weights and probabilities depend on validation with approved data.
Production use depends on the client’s data, privacy and SEO agreements.

Future opportunities

Validate embeddings and weights with approved data and measure relevance.
Connect the conversion feedback loop to ranking.
Extend demand forecasting with geography and seasonality.
Publish measured results when there is authorization and data.

Open the interactive algorithm tour

The tour walks through the concepts with interactive labs.