Skip to content

The platform

From your documents to a system in production.

Not a model handed over in a notebook — a maintained production system. Every step of the pipeline runs inside your jurisdiction, from data preparation to monitoring.

Pipeline

Seven steps, one jurisdiction.

01

Discovery & compliance scoping

We map the use case, data sources and the regulatory frame (FADP/nFADP, GDPR, EU AI Act), and define success metrics and guardrail requirements up front.

02

Private, in-jurisdiction data prep

Data is curated, cleaned and structured where it already lives. Nothing is shipped to an external service. PII handling is designed in from the start.

03

Fine-tuning on Apertus

We fine-tune Switzerland's open Apertus model on your corpus to teach your domain, terminology, formats and tasks — validated against a retrieval-only baseline, with full transparency over weights and data lineage.

04

Agent & RAG orchestration

Agents, retrieval-augmented generation and tool calls into your systems, so the model reasons over live knowledge and takes real actions.

05

Evaluation & guardrails

A use-case-specific eval set, safety and PII guardrails, and output constraints. We report quality against agreed metrics before go-live.

06

Sovereign deployment

Containerised deployment to your Swiss cloud, your own hardware, or confidential-compute VMs (AMD SEV-SNP / Intel TDX with remote attestation). Zero dependency on US LLM APIs.

07

Monitoring & iteration

Continuous monitoring of quality, drift and usage, with a maintenance cadence and a support agreement. The system improves as your data and needs evolve.

Architecture

Your data never leaves your jurisdiction.

A single trust boundary contains everything — data, model and orchestration. There is no round-trip to an external API.

  • Apertus core
    Fine-tuned Swiss open LLM — open weights, auditable, yours to run.
  • Agent orchestration
    Stateful agents, routing and tool use over your systems (LangChain + LangGraph).
  • RAG layer
    Retrieval over your knowledge bases, grounded and cited.
  • Guardrails & eval
    Safety, PII and quality checks, monitored in production.
  • Sovereign runtime
    On-prem, Swiss cloud, or confidential compute.

Model-agnostic by design: Apertus, Llama or Mistral — whichever wins on your evaluation set. The evaluation decides, not the logo.

Architecture: your data stays inside your jurisdiction. sysf.io fine-tunes Apertus and orchestrates agents, retrieval and tools, with no data sent to third-party US LLM APIs.
Third-party US LLM API
Your jurisdiction
  • 01
    Your data
    Documents, records, systems — never leaves
  • 02
    Fine-tuned Apertus
    Swiss open model, trained on your corpus
  • 03
    Agent orchestration
    Agents · RAG · tools · guardrails
  • 04
    Your product
    Copilots & workflows — yours to run

On-prem · Swiss cloud · Confidential compute — zero data egress

Compliance

Compliance is an architecture, not a checkbox.

Swiss FADP / nFADP

In-jurisdiction processing and clear data lineage keep personal data under Swiss law and control.

EU GDPR

Data minimisation, purpose limitation and residency handled by design for EU operations.

EU AI Act

Documented training data, transparency and guardrails support risk classification and obligations.

We deliver compliance and audit documentation alongside the system. This is not legal advice — we work with your compliance and legal teams to meet your obligations.

FAQ

Platform questions.

What is the technical stack?
A fine-tuned Apertus model at the core, orchestrated with LangChain and LangGraph for agents, retrieval-augmented generation (RAG) over your knowledge, and tool calling into your systems. We add evaluation harnesses and guardrails, and deploy via containers to your Swiss cloud, on-premise hardware, or a confidential-compute enclave.
Do you fine-tune, or just do RAG?
RAG first; fine-tuning where it earns its keep. Retrieval grounds answers in your current documents and is where most accuracy comes from. Fine-tuning teaches your domain's terminology, formats and task behaviour — it does not reliably memorise facts, so we never use it as a substitute for retrieval. Every Pilot ships with an ablation on your evaluation set — base model vs. RAG vs. fine-tune vs. both — so you see exactly what each component buys before you invest in production.
What does on-prem actually require?
Less than you'd think. A quantized, fine-tuned Apertus-8B runs at interactive speed on a single workstation-class GPU server (roughly 24–48 GB of VRAM) that lives in your existing rack. The 70B needs a multi-GPU node or a confidential-compute instance. We size the deployment during scoping and commit to throughput numbers — and a Pilot can run on hardware we bring, so you spend nothing on infrastructure before value is proven.
How do you handle evaluation and guardrails?
Every engagement includes a use-case-specific evaluation set and metrics agreed with you, plus guardrails for safety, PII handling and output constraints. We report quality against those metrics before go-live and monitor them continuously in production.
Where does it run?
Wherever your data governance requires — your Swiss or EU cloud tenancy, your own datacentre, or an air-gapped environment with confidential compute (AMD SEV-SNP or Intel TDX with remote attestation, so even the infrastructure operator sits outside the trust boundary). Nothing depends on external US LLM APIs.

See it on your use case.

Bring one high-value workflow. We'll show you the pipeline end-to-end — on your data, in your jurisdiction.

Book a call See how it works