Data company · Brazil
Urbyx is a data company. Since 2021 we have refined, reconciled and verified geographic and public information at national scale — first inside some of the world's largest technology products, now in a base of our own covering 52 million Brazilian addresses, in custom datasets for client systems, and in corpora built for the models being trained on this country.
What we do
Public data in Brazil is abundant and badly joined. Everything we do comes from the same craft — taking sources that were never designed to work together and making them agree well enough to decide on.
Our own geocoded record of 52 million Brazilian addresses, each carrying more than 170 attributes drawn from official sources and kept current as those sources publish. It is licensed as bulk extracts, as enrichment of your list, or through an API.
Entity resolution, name reconciliation, deduplication, verification against ground truth. This is the part that decides whether a dataset is usable, and it is the part most teams underestimate. We do it on our own base and on data our clients already hold.
Structured, labelled and verified datasets for training and evaluation. We spent three years doing this work for major technology companies before Urbyx existed, and we are returning to it — this time on a base we own and can document end to end.
Track record
For more than three years we refined cartographic data for one of the world's principal mapping platforms: verifying addresses, correcting geocoding, judging the quality of information that millions of people rely on every day to get somewhere. In parallel we worked on the evaluation and training of artificial intelligence systems for large technology companies, including language models now among the most widely used in the world.
That work taught us two things, and they still decide how everything here is built.
The value is never in having the file. It is in the reconciliation, the verification and the maintenance that make the file safe to join against something else.
Every attribute we ship traces back to a named source with a stated reference date. Where a column is sparse, the documentation says so and says why, rather than filling the gap with a model and calling it data.
The base
The base is the clearest thing we can show you, so here it is at full resolution. Most vendors describe a region and let you infer the address; we describe the address and let you aggregate upward. Field names below are the ones you receive.
Why this is hard
The public sources are good. What they do not do is agree on how an address is written, and nothing in Brazil issues a single key that joins them.
A street is spelled one way in the federal address register and another on every listing. A neighbourhood is called one thing by residents and delimited differently as a statistical unit. The same street name repeats across a city. Each of these is resolvable, and resolving them at national scale is years of work — which is what you are licensing as much as the attributes themselves.
Official statistics delimit neighbourhoods as a mapped geography. The national address register records the neighbourhood name as it is actually written on the address — hundreds of thousands of distinct names, informal ones included. Both are correct and they answer different questions. Connecting the name in use to the geography carrying the statistics is the join, and it is not a string comparison.
The federal address register and the commercial market rarely agree on how a street is written. Matching them is entity resolution, not string comparison, and we run it as a maintained process rather than a one-off clean.
A single street name can occur several times inside the same municipality, in different neighbourhoods, with no distinguishing identifier in the address itself. Disambiguating which one an address refers to is a routine step in our pipeline and a routine silent error in files that skip it.
Coverage
Every layer hangs off the same geocoded address record, so a client working on flood exposure and a client working on retail expansion are querying the same spine — and can be given a single file if they need both.
Physical exposure per address, not per postcode. Built for underwriting, portfolio screening and climate analytics in a country where flood and landslide losses are concentrated at street level.
How the country is actually built and who lives there, resolved to the smallest published statistical unit and carried down to the address.
Price and property signal for a market with no MLS and no public transaction register. We build the series ourselves and state plainly what each figure is.
For AI
Models that reason about Brazil are trained on whatever was scraped. That is fine until the task is an address — at which point the difference between data that exists and data that has been verified becomes the difference between a right answer and a confident wrong one.
We came into this from the service side: for three years the work was refining and judging other people's data — cartographic verification for a global mapping platform, evaluation and training for large language models. Good work, and someone else owned the asset at the end of it.
We are returning to that line now with a different footing. The corpus is built on a base we own, so we can document provenance end to end, state the fill rate of every field, regenerate a cut to a new specification, and licence it under terms that are ours to give.
If you are training or evaluating on Brazilian geography and the ground truth is the bottleneck, that is the conversation we want.
Delivery
There is no fixed package. You describe the decision you are trying to make; we tell you what the base can answer, then agree scope and price before any work starts.
A defined slice — a state, a set of municipalities, a radius, a risk band — delivered once or on a refresh schedule you set.
You send an address list or a portfolio; each row comes back with the layers you selected attached, matched against our resolution process.
Query by address, coordinate or radius from inside your own systems, for products that need the lookup live rather than batched.
Evaluation
An evaluation sample is available on request: real addresses across twelve municipalities spanning all five Brazilian regions, chosen to expose the range rather than flatter it. It ships with a data dictionary stating the measured fill rate of every column — including the sparse ones, with the reason each is sparse.
Who we work with
Pricing flood and landslide exposure on a book where two addresses on the same street can sit six metres apart in elevation.
Physical risk scoring for Brazilian assets, with a documented lineage back to federal sources rather than a modelled proxy.
Valuation context, portfolio screening and listing enrichment in a market with no public transaction register.
Site selection and territory planning against catchment demography that resolves below the municipality.
Verified Portuguese-language geospatial ground truth, and collection built to a specification you set.
Network planning and service-gap analysis using household-level supply, sanitation and drainage coverage.
Compliance
Our licensed products describe places, not people. They contain no names, no identification numbers, no contact details and no household-level records. Socio-economic attributes are published statistics aggregated by the national statistics agency at census-tract level, never observations of an individual household.
We consider these products outside the scope of personal data under Brazil's LGPD and the GDPR, and we apply minimisation and aggregation as standard regardless. Licence agreements prohibit any attempt to re-identify individuals from the data. The full notice is published at urbyx.com.br/data-privacy.
Contact
Describe what you are trying to answer and we will tell you what the base can and cannot do for it — including when the answer is that it cannot. You will hear from a person who knows the data, usually within one business day.