Skip to content

Kult Builds a Voice Gaming Companion on Gemini Flash Models and Google Cloud

London startup Kult has launched a voice companion that answers in-game questions, backed by a mix of Gemini models on Google Cloud. It now covers more than 100 games and is in alpha with first players.

3 min readBased on 1 source0 views
AI & ML

TL;DR

  • Kult is a London startup building a voice and second-screen companion that answers questions about the game a player is currently in.
  • It routes work across Gemini models: lightweight Flash models handle every conversational turn, while larger models do heavy reasoning once per game.
  • The product covers more than 100 games, adds new ones automatically, and is in alpha with first players.

The problem: answers that arrive too late

Getting help mid-game usually means alt-tabbing to a video or a forum thread and hoping the answer is still relevant when you find it. Kult wants the answer to come to the player instead. A player asks out loud, hears a short reply, and can check a second screen for more detail without leaving the game.

According to the Google Cloud Blog, three constraints decide whether this works: latency, accuracy and cost. Responses must be fast enough to keep up with play. Accuracy is harder, because general-purpose models have training cutoffs and can be months out of date, and they often get specific patches or builds wrong. Each game also has its own items and vocabulary. Cost matters because an always-on companion handles many turns per session.

Model routing instead of one model

Kult uses a range of models rather than a single one. The Google Cloud Blog describes the lightest Gemini Flash tier as the workhorse for high-volume, repetitive work, including the thousands of small calls made for each game added. Flash also runs every conversational turn: it works out what the player is asking and writes an answer grounded in Kult's knowledge of that game. The most advanced Gemini models handle the rare, heavy reasoning that happens once per game.

Kult reaches the models through the Gemini platform and the Gemini Developer API, with a routing layer that picks the backend for each model. The backend runs on Google Kubernetes Engine. Kult also uses a decision-only model it calls Jev, which returns a choice with a calibrated confidence score. Confident decisions are settled quickly, and uncertain ones escalate to a larger Gemini model. Each model sits behind a registry, so a new generation is benchmarked against reference answers written by people who know the games before it replaces an existing one.

Answers built from a knowledge base

Rather than relying on what a model remembers, Kult builds a structured knowledge base for each game automatically and keeps it current, so that a game changing with each patch is reflected in answers. The model's job is to reason over that knowledge. The company also tracks session context: the current game, recent actions and the previous question. A question such as "What is he weak to?" can be answered without restating the boss's name, because Kult already knows which fight is underway.

Every answer has two parts. The voice gives a short version, while a visual feed on the second screen shows images, video, maps, builds, interactive tools and links to guides. The feed is built with Google's open standard for agent-generated interfaces. The model does not write code. It picks components from a catalog that Kult designed, and the app renders them in its own style. Kult says this keeps the feed consistent and lets it add new kinds of visual answers without rebuilding the front end.

Kult states that it is also working on computer vision, so the companion can notice what is happening on screen without being asked, using Google's Live models for video understanding. That work is still ahead.

Why it matters

For developers, the case is a concrete example of a multi-model architecture in production. Routing frequent, cheap calls to a small model and reserving larger models for rare, complex tasks is a common pattern for keeping per-request cost manageable. Using a classifier with a confidence threshold to decide when to escalate is another reusable idea, as is a registry that lets teams benchmark a new model against reference answers before switching.

The approach also shows why grounding matters. Kult's bet is that a maintained, game-specific knowledge base beats relying on a model's memory, especially when content changes with patches. Teams building domain assistants can take the same lesson: keep the facts in a store you update, and let the model reason over them. Note that the company's own account is a vendor case study, and it has not published independent accuracy or latency figures.

Sources

This article was written by Blozhik from the reporting below. Read the originals for full detail.

GoThe Blueprint: How Kult built a gaming companion that actually knows your gameGoogle Cloud Blog
Was this useful?

Keep reading

Discussion · 0

Sign in to join the discussion. Liking does not need an account.Sign inRegister
Loading comments…
blozhikDeveloper news from across the web, explained. Every story links to its original sources.