The Missing Schema for the Physical World

A World Without Data#

Once upon a time, physical spaces had no data model. A football pitch was one hundred and five meters of grass. A warehouse was shelves and forklifts. A port was cranes and containers. The physical world existed (massive, complex, constantly in motion) but it was opaque to software. There was no way to ask it a question.

"Think about what we take for granted in the digital world," Suresh says, his voice carrying the cadence of someone who has delivered this thought experiment many times and still finds it remarkable. "Every digital interaction generates structured data. A click has coordinates, a timestamp, a user context. A transaction has an amount, a counterparty, a ledger entry. You can query any of it. You can aggregate it. You can build predictive models on it. The digital world is natively queryable."

He pauses, drawing a contrast that sits at the center of everything he has built.

"Now consider a football match. Twenty-two players, a ball, officials, a pitch with zones and lines and rules. Over ninety minutes, thousands of events unfold: passes, tackles, sprints, formations, set pieces. The amount of spatial information is extraordinary. And historically, none of it was captured in a way software could understand. Cameras recorded video. Humans watched the video. The spatial intelligence existed only in the minds of coaches and analysts who watched the footage."

A football pitch was not a database. It was just grass.

Cameras That Watched But Didn't Understand#

Every day, cameras watched physical spaces and produced pixels, billions of them, streaming at sixty frames per second, stored in video files that no software could natively query.

"The video contained everything," Suresh explains. "Position, speed, relationships, events, all there in the footage. But video is the wrong format for intelligence. You cannot ask a video file: which players were within passing distance of the ball carrier while moving into open space? You cannot query a video stream for how the formation evolved over the last thirty seconds. The information is there. The structure is not."

This was the fundamental mismatch. The physical world generates spatial information continuously and abundantly. The data infrastructure to capture, structure, and query that information did not exist.

"The physical world doesn't organize itself by rows in a spreadsheet," Suresh says. "It organizes itself by space. A person exists at a position. They have a speed and a direction. Their relationships with other people are fundamentally spatial: who is near whom, who is inside which zone, whose paths are about to cross. When you force spatial questions into traditional databases, you pay a performance penalty on every query. It's possible. It's also fundamentally the wrong tool for the job."

The Venue Graph#

One day, the idea crystallized into an architecture: the Venue Graph.

"The Venue Graph is a live digital twin of everything happening at a physical venue," Suresh says, and the way he says it (slowly, precisely) reveals that this definition was hard-won. "Every entity (player, ball, official, camera, zone) has a position, a speed, a direction, and a shape. Every relationship (who is marking whom, what formation is being played, who is in which zone) updates continuously. And the entire model is time-stamped. You can query what is happening right now, or replay any moment in history."

This was not just a detection log. Detection logs store "we saw something here at this time," isolated observations with no relationships. The Venue Graph stores understanding: entities with real-world physics, connected by relationships that evolve in real time.

"When you query the Venue Graph, you don't ask 'where is player seven?' You ask 'given how everyone is positioned right now, which passing lanes are open, which defensive zones are exposed, and how has the formation evolved over the last forty-five seconds?' These are real-time spatial questions over a constantly changing world. No existing database could answer them."

Building Arcflow#

Because no existing database could serve these queries, they built Arcflow.

"Arcflow is a high-performance engine purpose-built for understanding space and movement," Suresh says, with the pride of someone describing something they built from first principles. "Let me explain what that means concretely."

In Arcflow, space is the primary organizing principle. When you ask "find every player within ten meters of the ball whose path will cross the passing lane in the next two seconds," that question is answered instantly, because the engine was designed from the ground up to answer exactly these kinds of spatial questions. A traditional database would have to do that calculation the hard way, assembling the answer from thousands of individual lookups.

The relationship layer makes connections between entities first-class. Player A is marking player B. The ball is in zone C. Camera 3 can see region D. These relationships update continuously as the situation on the pitch evolves.

And all of this runs on the same AI chips at the venue that power the vision models. The entire pipeline, from what the cameras see to answering spatial questions, stays on-site, with zero cloud dependency. That means real-time answers with no internet lag.

Suresh Gohane

OZ Cortex / AI Stack Lead

AI Runtime & Cortex

“Every physical venue is a spatial database waiting to be queried. We just gave it a schema.”

From Data Layer to Platform#

Because Arcflow provides the spatial data foundation, the Venue Graph became the structured output layer that every downstream product builds on.

"This is where the architecture becomes strategically powerful," Suresh says, leaning forward. "The AI vision models output roughly seven thousand detection events per second across six cameras. Raw detections at that volume are noise. No downstream customer (not a broadcast director, not an analytics platform, not a security system) can use raw detections."

The Venue Graph transforms detections into understanding. Instead of "we see something at these pixel coordinates with 94% confidence," customers receive "player 7, home team, 34 meters from goal, sprinting northeast at full speed, marking opposing player 10." That is the difference between data and knowledge.

The Spatial API gives partners stable, reliable access to this knowledge. Downstream teams build on the Venue Graph with confidence that the data format is stable, the data is correct, and the queries are well-defined. They don't need to understand how multi-camera AI works. They ask the Venue Graph a question and receive a clear answer.

Structured Output for Every Product#

Until finally, the Venue Graph became the single source of truth, not just for individual venues, but for the emerging world model itself.

"The compounding happens at two levels," Suresh explains. "Within a venue, every match adds real-world data to the model. The system understands the venue more deeply over time: its lighting patterns, its wind effects on ball trajectories, its crowd movement dynamics. Across venues, the collective data builds a generalizable model of how physical spaces behave."

The data flywheel is transformative. Every match captured generates training data for predictive models, real-world sequences from conditions that no synthetic dataset can reproduce. Every match is a training example. The more matches, the better the predictions. The better the predictions, the more valuable every downstream application becomes.

Arcflow is going open-source because the data layer for physical-world intelligence should be a standard. "Our advantage is not the database engine," Suresh says. "It's the full vertical stack above it: the hardware, the AI models, the operational playbooks, the venue relationships. You can download Arcflow. You cannot download our custom hardware, our proprietary designs, or our deployment history. And every application built on Arcflow validates our thesis and expands the ecosystem we're best positioned to serve."

Every Space Becomes Queryable#

And ever since then, every venue becomes measurable. Every spatial relationship becomes data. The world model compiles.

"From perception to prediction," Suresh says, describing the trajectory with undisguised wonder. "Today, the Venue Graph tells you what is happening. The next step is telling you what will happen. Given how everyone is positioned right now (where they are, how fast they're moving, what formation they're in), what are the likely outcomes two seconds from now? Five? Ten?"

He pauses, and the academic precision gives way to something more expansive.

"And the beautiful thing, the thing that genuinely excites me, is that this isn't specific to football. The spatial primitives are universal. Entities in space, relationships between them, events unfolding in time. A warehouse. A port. A military installation. A hospital campus. The physics changes. The spatial primitives don't."

He returns to his foundational claim, the one he has tested against a PhD in robotics, years of production spatial systems, and every match the Venue Graph has processed.

"Space is the first primitive. Time is the second. Everything else (every event, every relationship, every prediction) is derived from those two dimensions. That's the thesis. And every venue we deploy, every match we capture, every query we answer makes that thesis harder to argue against."

A football pitch is no longer just grass. It is a spatial database with a schema, a query language, and a version history. And every pitch, every venue, every physical space in the world is next.