Under the hood: Building real-time race insights with Gemini and Google Cloud

05 Dec 2025 3 Mins

In Formula E, milliseconds don’t just decide the winner on the track; they decide whether we capture the fan’s attention at home.

With the GEN3 Evo race car, the racing is faster than ever, and the data generated is immense. The challenge? To take that torrent of telemetry and timing data and turn it into a human-readable narrative while the cars are still battling through the corner.

We recently launched a new Google Cloud and Formula E Strategy Agent built with Google Gemini to solve exactly this. By leveraging Gemini models, we are now able to provide race insights that appear on the broadcast in near real-time.

Here is a look under the hood at the architecture built to make it happen, all powered by Google Cloud Platform (GCP).

1. The Fuel: Data Ingestion

Everything starts with the raw feed. We process two primary high-fidelity sources of data:

Timing System: A data stream from our official timing provider with events such as lap times and driver positions
Telemetry System: A stream of each cars telemetry - such as battery status, speed and throttle traces

The ingestion process utilises a combination of Airflow, Cloud Scheduler and Cloud Run to pull data from these external APIs and streams. This serverless and automated process is tightly synced to our official race calendar.

2. The Nervous System: Pub/Sub and Storage

Once the data enters our Google Cloud environment, speed is critical.

The data is immediately forwarded to Pub/Sub, acting as our asynchronous messaging backbone. This allows both internal apps and external partners to subscribe to the feed with minimal latency.

3. The Transformation Layer

We utilize Cloud Run services to load data directly into AlloyDB. We chose AlloyDB here over our traditional BigQuery data lake because we needed a PostgreSQL-compatible engine capable of handling intense transactional loads with sub-millisecond latency.

4. The Brain: The Strategy Agent

The core intelligence of this system lives on a dedicated Compute Engine instance which hosts our "Strategy Agent."

This Agent is the bridge between raw numbers and race context. It connects directly to AlloyDB and continuously monitors the session. It isn't just watching a stream; it is evaluating the state of the race using a complex set of triggers and rules like "is a Safety Car deployed? Or "has the leader’s energy dropped below target?"

Relevant data is then returned for analysis by Gemini. Future development of the Agent will likely pass more raw data directly to Gemini to allow on the fly decisions to be made, or incorporating Machine Learning techniques.

5. The Generative Engine: Gemini 2.5 Flash

Once the Strategy Agent has packaged the relevant data points, it sends them asynchronously to Gemini 2.5 Flash.

While Gemini 3 has been released, the Flash variant is not yet publicly available; we plan to upgrade as soon as it is released. We chose the Flash model for its balance of reasoning capability and extreme speed. The AI analyses the data packet and generates a natural language insight—essentially turning a spreadsheet of lap times into a sentence like: "Vergne is overconsuming energy in Sector 2 to close the gap."

6. The Finish Line: Broadcast Integration

The final mile is where the digital meets the visual. The generated text insight is automatically passed to the broadcast graphics engine. The generated graphic is sent to the broadcast director for approval, before going live to millions of screens.

7. Observability

Because this is live TV, live monitoring and alerts are essential. We monitor the entire pipeline from end-to-end using Masthead Data for real-time observability, ensuring that if a data pipe clogs or a latency spike occurs, we know before the viewers do.

What's Next?

This architecture has laid the groundwork for a new era of sports broadcasting. We aren't just showing fans what is happening; with Google Cloud and Gemini, we are explaining why it matters, in real-time.