Welcome to The Blueprint, our ongoing series where we highlight how Google Cloud customers are tackling unique and common challenges across industries using the latest AI and cloud technologies. We hope to inspire others looking to innovate in their work.
Kult is a London-based startup building a companion experience for games. It sits alongside your session, understands the game you are playing and where you are in it, and answers by voice and on a second screen. The more you play with it, the better it knows how you like to play.
The challenge
Traditionally, getting help in the middle of a game meant alt-tabbing, opening a video, scrolling a forum, and hoping you find the answer before you forget why you looked. Information should come to the player instead. As gamers ourselves, our team wanted a way for someone to be able to ask a question out loud, hands on the keyboard or controller, hear an answer, and consult a second screen if needed, without interrupting their immersion in the game.
Three things decide whether that works: latency, accuracy and cost.
Latency: Games are time-sensitive. The answer has to be fast, and the interaction should feel like asking a friend who’s an expert on the game, sitting on the couch next to you and watching you play.
Accuracy: General-purpose models have training cutoffs. They can be months out of date and are often wrong about a specific game, patch, or build — and a wrong answer can be costly in the game and a waste of the player’s time. Every game also has its own items, systems, and vocabulary, so accuracy has to hold across hundreds of games. The right answer also depends on the player: The same question deserves a different answer 10 minutes into a game than a 100 hours in.
Cost: An always-on companion handles many turns per session, so every turn has to be inexpensive enough for a consumer product. Covering hundreds of games is only possible if adding one is cheap and takes no manual work.
The solution
Given these needs, we knew we would want a range of models to build Kult. Fortunately, Google’s Gemini models have some of the greatest variety in terms of balancing performance and especially cost, and with the Model Garden in Gemini platform, we have access to hundreds of models beyond Gemini. Gemini models have became the default family on both sides of the system: the live conversation, and the knowledge behind it, dynamically match the right model for every task.
The lightest Gemini Flash tier handles high-volume, repetitive work, where we make thousands of small calls for every game we add. Gemini Flash runs every conversational turn. It works out what the player is asking and writes an answer grounded in our knowledge of that game, fast enough to keep up with play. The latest and most advanced Gemini models handle the rare, heavy reasoning we do once per game.
Having that range on one API surface matters to a small team. The latest Gemini Flash models in particular sit at a point of price, speed, and quality that we have not found elsewhere, and they are the reason an always-on gaming companion works as a consumer product.
We reach the models through both the Gemini platform and the Gemini Developer API, with a routing layer that picks the backend for each model. Our backend runs on Google Kubernetes Engine.
Every answer comes in two parts. The voice gives the short version. At the same moment, a visual answer lands in a feed on the player’s second screen: images, embedded video, maps, builds, interactive tools made for that game, and links to the guides and creators behind the answer.
We build that feed with Google’s open standard for agent-generated interfaces. The model does not write code. It describes the interface it wants using a catalog of components we designed, and the app renders it natively in Kult’s own style. That keeps the feed safe and consistent, and lets us add new kinds of visual answers, and new devices, without rebuilding the front end.
The architecture
The outcome
Kult now covers more than 100 games, and adding a game is an automated process, not a manual project. That’s a key part of what lets a small team aim at any game a player picks up, including new releases, where most of the playing happens in the first weeks.
We are in alpha with our first players and opening up to more over the coming months. As we have grown, we have built a close working relationship with the Google Cloud team, with direct access to their engineers when we need it. That matters when every model generation raises the ceiling of what a gaming companion product can do.
This expertise, both within the models and our supportive partners, allows us to focus on our passion: building the features gamers want, faster and better than any possible competition.
The details
Five principles shape how Kult works.
Answers come from knowledge we build, not from what a model happens to remember. A game can change with every patch. For each game we cover, Kult builds a structured knowledge base automatically and keeps it current, and the model’s job is to reason over that knowledge.
The answer depends on who is asking. Kult keeps track of your session: the game, what you have been doing, and what you asked a moment ago. “What is he weak to?” needs no further explanation, because Kult already knows which boss you’re facing. Over time it learns from how you play, what you ask about and what you ignore, and shapes its answers around you. A player who wants the fastest route and one who wants to find everything should not get the same replies.
The player is never left in silence. Kult responds straight away and carries the conversation while the deeper reasoning runs behind it. The spoken answer is short, and the visual feed carries everything the voice cannot.
The smallest model that can do the job does the job. Much of the work in a system like ours is a decision, not a piece of writing. It needs to understand which game the query is about, does the question need a lookup, did the answer cover the point.
For decisions like these we use Jev, a decision-only model that returns a choice with a calibrated confidence score. Confident cases are settled in a fraction of a second at a fraction of the cost, and uncertain ones escalate to a larger Gemini model. Every model sits behind a registry, so when a new generation ships, we benchmark it against reference answers written by people who know the games, and swap it in if it wins.
Every question makes Kult better. Taken together, the questions players ask about a game show us where its knowledge is thin and what to build next. We are designing Kult so that this loop tightens with every session. The more people play, the richer the knowledge becomes for everyone. Every session teaches a player something, as well as the entire Kult system, and Kult helps them keep it. We think the companion that sits beside the game is only the start.
Next is computer vision. In the coming months, Kult will start to see what is happening in your game, so the companion can be aware of the moment without being asked. We’re using Google’s Live models as part of that work, for video understanding and more.
If you are a gamer, try it for yourself www.joinkult.gg
