Compliance-grade AI, built to hold up under scrutiny
A Swiss AI workspace for public administrations, clinics, schools and SMEs, where every byte of client data stays inside Switzerland.
- Client
- aineo GmbH
- Location
- Rapperswil-Jona, Switzerland
- Industry
- AI for the public sector, healthcare & SMEs
- Timeline
- 2024 to present
- Engagement
- End-to-end product build, ongoing
- Platforms
- Web, Microsoft Outlook, iOS
The challenge
aineo GmbH set out to help Swiss organisations adopt AI deliberately: municipalities, medical practices, schools, utilities and consultancies that wanted the productivity of large language models without handing their documents to a US cloud. The brief was a multi-model AI workspace that complies with the Swiss Federal Act on Data Protection (DSG), keeps data in Switzerland, and never trains on customer content.
Compliance was only half of it. The product had to feel instant in a conversational interface, survive rate limits and latency spikes from upstream model providers, and meet an enterprise uptime commitment. And it had to be usable by people with no AI background at all.
Our approach
We treated data residency as an architecture decision rather than a checkbox. Swiss-hosted open models (Qwen, Apertus, GPT-OSS, Mistral, DeepSeek) are the default. International providers (OpenAI, Anthropic, Gemini) are available, clearly labelled, and opt-in per organisation. Document embeddings, files and transcripts never leave Swiss infrastructure.
On the engineering side we introduced full TypeScript typing across the API layer and component library, replacing the untyped code we inherited, and shipped every new use case behind a feature flag so aineo could roll features out to one customer at a time.
What we built
A use-case driven chat workspace where users attach PDFs, Word and Excel files or images, dictate by voice, and organise conversations by topic. Answers stream token by token over Server-Sent Events from a FastAPI gateway that routes each request to the healthiest available model and fails over transparently when a provider stalls.
Around the chat we built document generation for Word and Excel, quiz and translation use cases, audio transcription with speaker separation, an Outlook add-in that checks tone and rewrites drafts in place, and a retrieval pipeline that chunks and embeds an organisation's own documents into pgvector inside Swiss AWS. Role-based access per organisation and full German, French and Italian localisation, down to validation errors, round it out.
What's inside
Multi-model routing with fallback
A provider router weighs live health, rate limits and latency, and redirects to a secondary model mid-conversation without dropping the stream.
Streaming responses
Tokens reach the browser over Server-Sent Events from an asynchronous FastAPI pipeline, so answers start appearing in well under a second.
Document-grounded answers
Celery workers parse, chunk and embed uploads into PostgreSQL with pgvector. Retrieval is tenant-scoped and stays inside Swiss AWS.
Outlook writing assistant
An add-in that reads the draft you are writing, checks the tone, shortens or rewrites passages, and never makes you leave your inbox.
Transcription and document generation
Recordings in Swiss German or High German become searchable minutes with speakers, timestamps and action items. Chats turn into Word and Excel files.
Built for organisations, in three languages
Role-based access per organisation, persisted user preferences, and a UI localised in German, French and Italian.
The results
- Gemeinde Stäfa
- rappjmed
- Akademie St.Gallen
- EWJR
- AMK Management
- Schweizer Bildungszentrum
The Byte Whiz built our AI platform end to end: a DSG-compliant, multi-model dashboard that keeps our clients' data inside Switzerland. They turned a demanding compliance brief into a product we're proud to put our name on.
Michael GüntenspergerCo-founder & Owner, aineoMore work
Have a problem like this one?
Tell us where you're going. We'll tell you honestly how to get there, and what it will take.