Built on Google Cloud
AI is the core of what Kolosei ships, not a bolt-on. Two products use Google Cloud in production today — and they do it under the user's own key, so we never hold a user's credentials or content on our servers.
Built on Google Cloud
- Gemini API
- Cloud Text-to-Speech
- Cloud Translation v3
In production today
Each of these is a real dependency in shipping code — the exact client library is named below.
| Google Cloud service | Where | What it does |
|---|---|---|
Gemini API@google/generative-ai | LinguistPro | Morphological analysis, translation assistance and text explanation for Hebrew. Runs under a per-user API key stored in the browser only — zero server-side retention. |
Cloud Text-to-Speech@google-cloud/text-to-speech | LinguistPro | Per-line Hebrew audio with niqqud-aware pronunciation, so learners hear any text they paste in. |
Cloud Translation v3@google-cloud/translate · google-cloud-translate | LinguistPro & HDLE Premium | Neural translation of Hebrew source text. The official v3 client, with budget guards and usage tracking in the desktop product. |
Planned / in evaluation
Our second vertical — EgorGenom, rare-disease genomics — is where we expect to scale on Google Cloud next. This is explicitly forward-looking, not yet in production.
| Service | Vertical | Intended use |
|---|---|---|
| Vertex AI | EgorGenom · LinguistPro | Phenotype-to-gene scoring models; fine-tuned Hebrew morphology / translation models. |
| BigQuery | EgorGenom | Querying population-frequency datasets (gnomAD) at scale instead of static local dumps. |
| Cloud Storage | EgorGenom | De-identified genomic data and long-read archives. |
| Document AI | EgorGenom | Parsing clinical reports and lab PDFs that are currently handled manually. |
Why Google Cloud
Two reasons, specifically. First, model quality for Hebrew and multilingual work: Gemini handles Hebrew morphology, mixed RTL/LTR text and Hebrew↔Russian/English translation better than the alternatives we tested — which matters when the whole product is about understanding a morphologically rich language. Second, our architecture is offline-first and only calls the cloud for heavy compute (speech synthesis, neural translation, model inference). That keeps user data local by default and makes Google Cloud the natural place to run exactly the workloads that genuinely need a frontier model — and the place we want to grow as usage and the genomics vertical scale.
The privacy design
LinguistPro keeps a user's texts, audio, progress and notes in browser-local storage (OPFS + SQLite WASM). When the cloud is needed, the user supplies their own Google Cloud key, held in the browser only — so Kolosei never stores user credentials or content server-side, and the user's Google Cloud usage is governed by their own account and Google's terms. HDLE Premium follows the same principle on the desktop, encrypting credentials at rest with AES-256-GCM in the OS keyring.