Gemini, and the jobs it wins for us
Technology
Gemini is native to Google's own stack: Vertex AI, Firebase, the Agent Development Kit, and the generative media models, Imagen and Veo, all live next to it. When a job runs on GCP, needs image or video generation alongside text, or has to hold a lot of context across a long pipeline, Gemini is the model already in the room.
We run Gemini 2.5 Pro and Flash across a nine agent pipeline on Vertex AI, we were given special access to Gemini 1.5 Pro to rewrite legal literature at different reading levels, and we've put Gemini, Imagen, and Veo in the same product to generate playable game content. Model selection is a decision made per job, and on Google's own infrastructure, Gemini is usually the right one.
Why
Where Gemini wins
- Native to Google's own infrastructure
When a job runs on Vertex AI, in Firebase, or through Google's Agent Development Kit, Gemini is already in the room. We don't reach for it out of habit. We reach for it because the model, the compute, and the deployment target sit inside the same account, with no integration tax at the boundary.
- Nine agents, tiered to control cost
For a developer platform serving 800 to 1,000 daily users, we built a nine agent pipeline on Gemini 2.5 Pro and Flash. Moving evaluation and refinement onto Flash cut latency about 70 percent, and routing only the relevant examples into each prompt cut token cost per request about 73 percent, with no drop in output quality.
- Imagen and Veo, in the same build
For Google I/O 2025 we built Living Canvas with Google: a playable web game where a player's sketch becomes a working game object. Gemini generates the scene, Imagen and Veo turn drawings into objects and actions, and Angular and PhaserJS handle the state and physics underneath.
- One source, rewritten for every reader
Harvard's NeuroLaw Library gave us special access to Gemini 1.5 Pro to adapt dense neuroscience literature into multiple reading levels from a single source: a comprehension slider instead of one fixed translation. That's a long context, instruction following job Gemini was well suited to.
