Work · IT Solutions
Rafael
In progressRafael is the AI gateway we build for the Crypton ecosystem: one API that chooses the right model, holds live voice conversations and acts only with a person’s approval.
- Role
- Own product
- Period
- 2026
- Outcome
- AI gateway in production
07 · Rafael
Intelligence under human control.
Rafael is the AI gateway we build for the Crypton ecosystem: one API that chooses the right model, holds live voice conversations and acts only with a person’s approval.
Challenge
The problem.
Every product that adds AI ends up rebuilding the same parts: model access, fallbacks, voice, tools, quotas and logs. Each copy behaves slightly differently and has to be maintained on its own.
Crypton needed intelligence that stays accountable: an assistant that can take real actions inside the messenger, but never more than the person could do themselves, and never without their consent.
Approach
How we approached it.
01
One gateway, many models
Applications call one OpenAI-compatible API. Rafael picks the model for each request and reroutes quietly when a provider is unavailable.
02
People stay in control
Anything that writes or sends goes through the same four steps: preview, approval, exactly-once execution and a log entry.
03
Clear separation of roles
Crypton is responsible for people and data, Rafael for intelligence. Apps never call Rafael directly: rights, quotas and consent are checked in the Crypton backend first.
04
Proven in public
Rafael answers visitors of this website in chat and by voice. The most honest demonstration is the product working where anyone can try it.
What we built
What we delivered.
Text assistant
Streaming answers, image understanding, files and tool calls through an OpenAI-compatible interface.
Live voice dialogue
Real-time conversation with interruption, transcription and a short summary. Rafael recognises the language of the person speaking.
Agents with approval
Agents plan a task, show what they are about to do and wait for approval. Every run can be tested first, is logged and can be replayed.
AI secretary in Crypton
When the user is busy, Rafael takes the call, notes the message and agrees a time to call back.
Website assistant
On alpineascent.ch Rafael answers in twelve languages, opens the right page, switches the language and prepares an enquiry for our team.
Architecture
System at a glance.
Four layers: Crypton apps on iOS, Android and the web; the Crypton backend for accounts, rights, quotas and approvals; Rafael for model routing, voice, agents, files and knowledge; and the models themselves, from cloud providers to our own local models.
Requests whose data must not leave our infrastructure run on local models on our own servers.
On the website, every action is validated on the server and limited to this site, and the voice relay keeps provider keys away from the browser.
Numbers & facts
At a glance.
115
Gateway API methods
446
Assistant actions in Crypton
12
Languages on this website
In production, in active development
Status
What we learned
What we take forward.
- 01
An assistant earns trust through what it will not do: the approval step is part of the product, not friction.
- 02
Model choice belongs in the gateway, not in each app. Providers change faster than products.
- 03
A voice assistant has to show that it is thinking. Silence feels like a failure, even when the answer is on its way.
Next steps
What’s next.
- Rafael in more Crypton apps and partner services.
- More local models for workloads whose data has to stay in our own infrastructure.
- Deeper knowledge of projects and services, so agents can plan with real context.
Continue the conversation