Building a brain for Earth

We pride ourselves on the most extensive data collection in the industry. We find creative and versatile methods to source data, including from drone-imagery, car-cameras, public data, private data, user-crowdsourced data, and partner data providers.

Powered by our Data EnzymesData EnzymesJust as your stomach breaks down food into the nutrients your body can actually use, our data digestion pipeline takes raw, messy, dormant inputs and turns them into clean, structured, well-labeled data the model can learn from. technology, we’re able to expand our data collection to capture previously-untapped sources. Using our data digestion engine, we’re achieving the cheapest P/POI.P/POI — Price per Point of InterestThe cost to capture a single geospatial data point. Lower P/POI means richer, more affordable training data, a key economic metric for any geospatial foundation model. Data Enzymes allow us to turn dormant data usable, and to provide lower-cost enriched data sets to a wider customer base.

Data Collection and Fusion are the most important steps in our architecture. That’s why we’re building automatic and accurate data filteringData filteringThe process of selecting, refining, and removing unwanted information from a dataset, such as noise, duplicates, or irrelevant variables. Additionally, scoring each data point for confidence and cross-validating it against other sources. and labelling.Automatic labellingThe process of tagging raw, unstructured geospatial data with structured metadata so a model can reason about it. Our pipeline does this without human annotation.

Data scarcity and data quality, the Pinnacle Data Problem,The Pinnacle Data ProblemThere’s not enough high-quality and well-labeled structured data currently available for a truly general-applicable Large Geospatial Model. Insufficient or lacking-in-quality training data is a pinnacle bottleneck to foundation models today. The data is out there but currently hard to capture. is one of the hardest parts about building the LGM. To solve this we have built distinct methods to universally digest data. With Data Enzymes,Data EnzymesJust as your stomach breaks down food into the nutrients your body can actually use, our data digestion pipeline takes raw, messy, dormant inputs and turns them into clean, structured, well-labeled data the model can learn from. we’re able to fuse data together for our model to train on and reason over. Cheaper to harvest → more data → smarter model.

We care about Ground Truths.Ground truthA verified, real-world observation confirmed at a specific X, Y, Z coordinate and time. We vet our data with accredited organizations, as well as conduct internal validation audits,Internal validation auditsMethodology for internal validation audits are always developing. Currently, we perform cross-validation against proxy variables as well as probabilistic consistency checks, to assess whether the model’s predictions are statistically meaningful. In addition, we routinely conduct proprietary ground surveying to validate data with certainty. to ensure each attribute is truthful at a given X,Y,Z point.

Our core reasoning architecture considers temporal data,Temporal dataObservations carrying a timestamp. For example, a street’s road condition in 2012 and the same street today are separate observations; keeping both allows a model to learn an attribute timeline, which helps it reason about how a place has been changing, and its future prediction. and sifts through vast amounts of aggregated geospatial data, including anthropologic data.

Our model includes our concept of Earth Recipes,Earth RecipesA recipe is a set of rules for how something comes to be. Copper deposits, for instance, form where particular geological conditions coincide. Rather than asking where its training data happened to label copper, the model asks what the recipe for copper is, which attributes it depends on, and where on Earth those attributes line up, inferring a new situation from rules and reality. which are fed to expert-agentsExpert-agentsExpert agents in AI are autonomous, specialized entities within the model, designed to execute complex tasks, execute decisions, and use external tools within a specific domain. They combine domain-specific Earth Recipes knowledge with goal-driven reasoning. They work together in parallel and combine to reach conclusions. that collaborate to learn and reason. MagellanMagellanMagellan is the name for the flagship Large Geospatial Models currently in development by Columbus Earth. continuously learns and creates new patterns through a relational architecture.

One model, innumerable applications.

Columbus ProMore Soon