Question 1

What is Galya?

Accepted Answer

What is Galya?

Galya is taste infrastructure. We sit underneath AI products and agents and give them the one thing they're missing: a structured understanding of who their users are.

Right now, if you ask an AI to recommend something personal, it either guesses, or it asks you to explain yourself. That's because preference isn't in the model, it has to be fed in. Galya builds that feed. We index your catalog for meaning and track how users engage with it, then turn those signals into context any model can use, at inference, in real time, without stored profiles.

A simple example: a user browses a travel platform. They linger on boutique coastal properties. They skip anything corporate. They save three places that all share the same quiet, textured aesthetic. Galya reads those signals, structures them, and when your agent needs to recommend a hotel, it already knows.

What can Galya actually do for my product?

What does "taste infrastructure" mean?

How does Galya compute taste?

What signals does Galya use?

What does Galya index about content?

Where does Galya get its data?

How can I access Galya's model and reasoning capabilities?

What's the difference between Galya's model and larger LLMs?

What problem does Galya solve?

Does Galya store user profiles or personal data?

How is Galya different from a recommendation engine?

How much does Galya cost?

Question 2

What can Galya actually do for my product?

Accepted Answer

Galya plugs into your product and delivers three things: Personalized discovery. Search and recommendations rerank against what each user actually responds to. A minimalist user and a maximalist user get different top results from the same catalog. Customer intelligence. A dashboard that shows you who your users really are: clustered by what they engage with, not what they say. "Users who respond to understated luxury across travel, fashion, and interiors" is more useful than "women 25–34." Agent context. Structured taste context injected at inference so your agents act with preference awareness from interaction one. No reprompting and no generic fallback.

Question 3

What does "taste infrastructure" mean?

Accepted Answer

Taste isn't a list of favorites. It's a pattern, consistent, cross-domain, and mostly invisible to the person who has it. The same aesthetic that draws someone to a particular café shows up in the furniture they save, the clothes they buy, and the travel content they linger on. It's the same signal. Just expressed differently depending on context. Infrastructure is what makes a signal persistent and usable across systems. Galya structures taste the same way Stripe structures payments, so any product can plug in and act on it, without rebuilding the underlying logic.

Question 4

How does Galya compute taste?

Accepted Answer

Taste computation starts with a joint embedding architecture. This means we take two fundamentally different things, a person and a piece of content, and translate them both into the same geometric space. On the content side, vision-language models extract deep semantic attributes from every asset: aesthetic quality, mood, emotional tone, compositional style. Those attributes get encoded as high-dimensional vectors: a precise coordinate for what something is, not just what it's labeled as. On the user side, behavioral signals (dwell time, scroll depth, saves, skips, return visits) get encoded against those same content vectors. The result is a taste embedding: a coordinate that represents where a user sits in that same space, derived entirely from how they responded to meaning, not from anything they said about themselves. At query time, Galya measures the distance between a user's taste embedding and the vectors in your catalog. The closest matches aren't just what they clicked most, they're what fits, aesthetically and semantically. That match gets returned as structured natural-language context that an agent or model can act on directly.

Question 5

What signals does Galya use?

Accepted Answer

Galya reads two types of signals: what your content means, and how your users respond to it. On the content side, vision-language models extract semantic and aesthetic attributes from every asset: mood, tone, style, composition, intent. On the behavior side, Galya tracks how users engage: dwell time, scroll depth, saves, skips, and return visits.

Question 6

What does Galya index about content?

Accepted Answer

Most systems index content by what it is: category, price, tags, keywords. Galya indexes it for what it communicates. Using vision-language models, every asset in your catalog gets analyzed for the attributes that actually drive preference:

Question 7

Where does Galya get its data?

Accepted Answer

Galya uses two inputs from your product: your catalog and your users' behavioral signals. That's all that ever touches your environment. Separately, Galya maintains its own proprietary training dataset, collected independently and never connected to any client product. It's used exclusively to calibrate the cross-domain Taste Graph: the layer that understands how a user's aesthetic preferences in one category reliably predict their preferences in another. Your data stays yours. Galya's training data stays ours. The output is a taste signal precise enough to work across any domain you operate in.

Question 8

How can I access Galya's model and reasoning capabilities?

Accepted Answer

There are three ways to bring Galya into a product. You can query it from your backend: indexing content, searching against the Taste Graph, and pulling structured taste context into your application. You can instrument your frontend: tagging your UI so behavioral signals flow into the index automatically, without manual event tracking. Or you can wire it into an LLM agent: Galya's methods are available as function-calling tools, so any model that supports tool use can query taste context mid-reasoning. Most products end up using more than one.

Question 9

What's the difference between Galya's model and larger LLMs?

Accepted Answer

LLMs are built to reason. They're not built to know who they're reasoning for. A large language model has no persistent sense of what a specific person responds to. It can infer preference from what you tell it in a prompt, but that puts the burden on the user to articulate something they often can't. Ask someone why they love a particular hotel or a particular jacket and they'll struggle to explain it. The signal is in their behavior, not their words. That's the gap Galya fills. We don't replace the model, we give it something real to reason on. Structured taste context, computed from how users engage with content, delivered at inference. The LLM stays the reasoning engine. Galya is the preference layer it was missing.

Question 10

What problem does Galya solve?

Accepted Answer

The problem isn't that AI products don't try to personalize. It's that the mechanism they rely on, memorization, was never built for it. LLMs learn from patterns at scale. That makes them extraordinarily capable at many things. But preference is personal, contextual, and often invisible to the person who has it. It doesn't live in aggregate patterns. It lives in the specific, accumulated signals of how one person responds to content over time. Memorization can't capture that. Galya does.

Question 11

Does Galya store user profiles or personal data?

Accepted Answer

Galya is designed to be fully compliant with GDPR, SOC 2, CCPA, and equivalent data protection regulations, by architecture, from the beginning. No personally identifiable information is ever required. Galya doesn't build profiles on your users or store behavioral history against an identity. Taste is computed from the relationship between anonymous behavioral signals and indexed content, at inference, in real time.

Question 12

How is Galya different from a recommendation engine?

Accepted Answer

Galya is infrastructure. A recommendation engine is a product. A recommendation engine takes your catalog and surfaces what's most likely to convert, based on popularity, purchase history, and collaborative filtering. It's a closed system built to produce a specific output: a ranked list. Galya doesn't produce recommendations. It produces taste context: a structured, portable understanding of what a user responds to that any model, agent, or system can reason on. A recommendation engine consumes that context. So does a search system. So does an agent. So does anything else in your stack that needs to understand a user. One sits on top. The other sits underneath.

Question 13

How much does Galya cost?

Accepted Answer

Galya is designed to scale as you scale. We offer pricing packages tailored to whatever stage you're at, whether you're just starting to build or scaling an enterprise product. Book a call to find the right package for you.

Frequently asked questions.

Ready to give your AI taste?