Security & confidentiality

Recapro’s security, with no blind spots

AI inference (transcription and summary) is 100% self-hosted in France, and no third party sees the content of your meetings. This page details our architecture, our French hosting providers, our sovereign processing chain and our compliance posture.

Our hosting The on-premise option

The essentials, in one answer

Recapro produces the transcription and summary of your meetings using open-weight models running on its own servers, in France. No third-party language or transcription model is involved, and the content of your meetings is never shown in clear text to any third party. The Cloud offering is hosted on French sovereign infrastructure, with no link subject to US law in the processing chain.

Inference 100% self-hosted in France: transcription (whisper.cpp), summary (vLLM) and speaker separation (pyannote) all run on a Jetson AGX Thor machine that we operate.
No third-party language model or transcription engine: the content of your meetings is never sent to an external AI service such as OpenAI, Anthropic or Mistral.
French sovereign Cloud: hosted on French sovereign infrastructure (France-Nuage in Vendée, Scaleway in Paris), with no link subject to US law in the processing chain.
Encryption at rest and in transit: service data is encrypted both in storage and during transport.
Compliance: Recapro is HDS-ready, with no HDS or ISO 27001 certification claimed to date.
On-premise deployment available: the whole chain runs on a machine installed at your site, with an air-gapped mode possible.

Inference fully self-hosted, in France

The moment a model reads your data in clear text is the sensitive point of any AI transcription. At Recapro, that moment takes place on our own hardware, with no intermediary.

01

Transcription: whisper.cpp

Speech-to-text conversion is handled by whisper.cpp, run locally. No third-party transcription service is called.

02

Summary: vLLM

Summaries are generated by an open-weight language model served via vLLM on our infrastructure. No call to OpenAI, Anthropic or Mistral.

03

Speaker separation: pyannote

Identifying who is speaking (diarization) is handled by pyannote, also locally.

04

Hardware: Jetson AGX Thor

The entire inference runs on an NVIDIA Jetson AGX Thor machine that we operate in France.

AI inference is 100% self-hosted in France, and no third party sees the content of your meetings.

Hosting and datacenters

Service data is hosted in France and in the European Union. Here is where, and with whom.

Vendée (France): France-Nuage

Hosting infrastructure and fleet observability, operated in Vendée by France-Nuage.

Paris (France): Scaleway

The Kubernetes compute of the SaaS service runs on Scaleway’s fr-par region.

Scaleway publicly claims the ISO/IEC 27001 and HDS certifications for the scope of its own infrastructure. These certifications cover the hosting provider, not the Recapro application running on it.

Source: Scaleway, “Security and resilience” (accessed 28 July 2026).

Certifications and compliance: the real state of things

We prefer to state precisely what we have and what we do not, rather than let anyone assume certifications that do not exist.

Recapro

HDS-ready

Our architecture is designed to be compatible with health-data hosting. Recapro is not HDS-certified to date, and we do not claim it.

France-Nuage

Not certified

As of mid-2026, France-Nuage holds neither ISO/IEC 27001 nor HDS certification.

SecNumCloud

No qualification

Recapro holds no SecNumCloud qualification, and no process is underway to date.

Scaleway

ISO 27001 + HDS (host scope)

Scaleway publicly claims ISO/IEC 27001 and HDS for its own hosting scope. These certifications do not cover the Recapro application.

HDS-ready describes an architectural compatibility, not a certification. We will update this page as soon as a certification is actually obtained.

The on-premise option: everything stays on your machine

In an on-premise deployment, Recapro is installed on a machine you control. The whole application chain and the inference run on it.

01

Everything on a single machine

Database, API, workers, sync (PowerSync), summary engine (vLLM) and transcription (whisper) run on a 128 GB NVIDIA Jetson AGX Thor installed at your site.

02

No public inbound port

The only inbound access stays on your local network: Kong (8000), PowerSync (8080), Whisper (8001) and vLLM (8002). No port is exposed publicly.

03

Limited, documented outbound traffic

In standard mode (not air-gapped), the machine only emits heartbeats to api.recapro.ai, telemetry to France-Nuage (via an authenticated access) and an outbound administration tunnel. On first boot, it pulls images from GitLab and models from HuggingFace.

04

No customer data leaves the machine

The telemetry sent by the fleet agent is limited to machine metrics: CPU, memory, disk, uptime. The content of your meetings never leaves the device.

05

Air-gapped mode possible

Recapro can run fully offline, with no outbound connection at all.

06

License kill-switch

The license drives activation: a suspended license stops the inference containers.

To compare deployment models, see Cloud vs On-Premise and the sovereignty & AI inference explainer.

Frequently asked questions

01

Does Recapro send the content of my meetings to a third-party AI service?

No. Transcription and summary are produced by open-weight models running on our own servers, in France. No meeting excerpt is sent to OpenAI, Anthropic, Mistral or any other third-party model.

02

Is Recapro HDS-certified?

No. Recapro is HDS-ready: its architecture is designed to be compatible with health-data hosting, but it is not HDS-certified to date, and we do not claim it.

03

Is Recapro ISO 27001-certified or SecNumCloud-qualified?

No. Recapro claims no ISO/IEC 27001 certification and holds no SecNumCloud qualification, nor any process underway. Scaleway, our compute host, claims ISO 27001 and HDS on its own scope, which does not cover the Recapro application.

04

Where is the content of my meetings processed?

AI inference (transcription and summary) stays on our servers in France: no third party ever sees the content of your meetings in clear text. The Cloud offering is hosted on French sovereign infrastructure, with no link subject to US law in the processing chain. On-premise, everything stays with you: the content never leaves your machine.

05

Where is my data hosted?

The service compute is hosted with Scaleway in France (fr-par region), observability and part of the hosting with France-Nuage in Vendée. The database and audio storage rely on French sovereign infrastructure, with no link subject to US law.

06

In on-premise mode, what data leaves the machine?

In standard mode, only heartbeats, license telemetry and machine metrics (CPU, memory, disk, uptime) are sent. The content of your meetings never leaves the device. An air-gapped mode, with no outbound connection at all, is also available.