Your LLM. Your servers. Your data.
Sovereign LLM inference for e-commerce merchants. Open-weight models, GPU servers in European datacenters, OpenAI-compatible endpoints. Your prompts, catalog and customer data never leave the tenant we operate for you.
| Property | Value | Note |
|---|---|---|
| data.residency | 100 % EU | OVH, Scaleway, Hetzner, UpCloud |
| training.on.your.data | 0 | by default, zero retention |
| batch.throughput | 10,000+ | prompts per batch via Batcher.ai |
| api.compat | OpenAI | drop-in base_url switch |
> 01 / ARCHITECTURE
Sovereign inference, end to end.
Three layers, one operator. Your GPU servers in EU datacenters, open-weight models we deploy and update, a secured API consumed by your applications. No third party in the middle, no token leakage to a US provider.
Dedicated GPU instances, NVIDIA and AMD, provisioned at OVH, Scaleway, Hetzner or UpCloud. Hardened Debian, isolated tenant, monitored 24/7 by Fast-Mage sysadmins. No resources shared across clients.
Llama, Qwen, DeepSeek, GLM, Kimi, gpt-oss, Codestral, plus Qwen-Embedding for RAG and FLUX for image, and other families on demand. We benchmark, quantize (GGUF, AWQ), and track new releases. Fine-tunes on your data stay on your tenant, full ownership and exit on request.
OpenAI-compatible endpoints (chat, embeddings, vision), token-based auth, usage monitoring, rate limits. Open-WebUI included for your team's collaborative prompt work. Drop-in replacement for an existing OpenAI integration.
> 02 / WHY SOVEREIGN
What changes vs. OpenAI or Anthropic.
Calling a US API from a French or European e-commerce stack means your catalog, customer messages, internal documents and prompts cross the Atlantic, fall under US jurisdiction (CLOUD Act), and contribute, by default, to training datasets you do not control.
| Criterion | Public US APIs | Fast-Mage private AI |
|---|---|---|
| Data residency | US datacenters, transit through US backbone | EU datacenters, EU-operated |
| Legal jurisdiction | US (CLOUD Act, FISA 702) | EU, GDPR-aligned |
| Training on your inputs | Opt-out, varies by plan and ToS update | Zero, by default |
| Model ownership | Closed weights, vendor lock-in | Open weights, your fine-tunes |
| Pricing model | Per token, unpredictable at scale | Per GPU, predictable monthly |
| Trade-secret exposure | Prompts and outputs leave your perimeter | Stays within your tenant |
| API contract | OpenAI / Anthropic native | OpenAI-compatible, drop-in |
> 03 / E-COMMERCE USE CASES
Built for merchants, not for demos.
Concrete workloads we deploy in production on Magento, WooCommerce, PrestaShop, Sylius and Akeneo stacks. Each pipeline is reviewed by a human engineer before going live.
Product descriptions, SEO metadata, translations across 5 to 20 locales, image alt text. Batcher.ai orchestrates 10,000+ prompts per run against your private inference cluster. Output is staged, diffed against the previous version, and pushed to your PIM or CMS on validation.
Ingestion of product catalog, technical PDFs, internal wiki, past support tickets. Customer-facing assistants and internal copilots query a private vector index, with strict per-tenant isolation. Re-indexing is webhook-triggered on catalog updates.
Pre-filter incoming customer reviews, contact messages and Q&A submissions. Classification on toxicity, off-topic, spam and language. The model never sees content that would otherwise leave your perimeter to a third-party moderation service.
Sales, support and merchandising teams query an Open-WebUI workspace connected to your private models, with shared prompt libraries. Multi-agent workflows (inventory checks, CRM enrichment, SEO drafts) via CrewAI or n8n integrate over the same OpenAI-compatible API.
> 04 / AI AGENT STUDIO
Agents that act on your business.
Beyond chat and batch, we deploy AI agents wired into your systems (catalog, CRM, helpdesk) that run recurring tasks under supervision. Built with your teams, hosted on your private infrastructure, never a black box.
Ready-to-use agent library for recurring e-commerce tasks: inventory sync and checks, CRM enrichment, SEO metadata generation and audit. Connected to Magento, Akeneo or your PIM over the API.
Compose custom workflows where several agents coordinate: retrieve, reason, act, validate. CrewAI and LangChain-compatible orchestration, run on your private models.
Every agent is traced: triggered action, source, token cost, latency. Drift alerting, with a human validation loop on sensitive actions.
Agents run on your Fast-Mage AI hosting, on dedicated EU GPUs. Your catalog and customer data never leave our infrastructure.
> 05 / WHAT YOU GET
A turnkey private AI bundle.
Inference cluster.
- + Dedicated GPU servers (L40S, A100, H100)
- + Load balancer across GPU instances
- + Daily model updates, GGUF / AWQ conversions
- + Hardened Debian, 24/7 monitoring
- + OpenAI-compatible API gateway
Batcher.ai included.
- + Spreadsheet interface for bulk prompts
- + 10,000+ prompts per asynchronous run
- + Multi-provider routing (private + third-party)
- + Recipe library, e-commerce templates
- + Free core access for hosting clients
Team and agents.
- + Open-WebUI for collaborative prompting
- + Shared prompt libraries across teams
- + CrewAI and n8n integration patterns
- + Deep research API (private beta)
- + Usage dashboards per team and per key
06 / FREQUENTLY ASKED
Frequently asked questions.
Five key points.
Answers reviewed by our engineers. If your question is missing, just write to us, quick reply, no sales script.
Q01 Where exactly does the inference run?
+
Where exactly does the inference run?
On dedicated GPU servers hosted in European datacenters (OVH, Scaleway, Hetzner, UpCloud). Your prompts, embeddings and responses stay on infrastructure under our operational control. No traffic is routed to OpenAI, Anthropic or any US cloud provider unless you explicitly enable a hybrid route.
Q02 Is your stack compatible with my existing OpenAI integration?
+
Is your stack compatible with my existing OpenAI integration?
Yes. We expose OpenAI-compatible endpoints (chat completions, embeddings). Switching your SDK base_url and API key is usually the only change required. Existing tooling like LangChain, LlamaIndex, n8n, CrewAI or your in-house clients keeps working.
Q03 Which models do you deploy?
+
Which models do you deploy?
Open-weight models selected with you. For chat and reasoning: Llama, Qwen, DeepSeek, GLM, Kimi, gpt-oss. For code: Qwen Coder, Codestral, Devstral. For vision: Qwen-VL, GLM vision, Llama multimodal. For RAG embeddings: Qwen-Embedding. For image diffusion: FLUX, Stable Diffusion. Other families on demand. We handle quantization (GGUF, AWQ), benchmarking on your GPU footprint, and continuous tracking of new releases. Closed proprietary models stay possible via a hybrid setup if a specific task requires it.
Q04 How does RAG on our catalog and support docs work?
+
How does RAG on our catalog and support docs work?
We ingest your product catalog, PIM exports, support tickets, internal wiki and PDFs. Content is chunked, embedded and stored in a private vector index that we operate. Queries combine retrieval with the LLM, with strict tenant isolation. Re-indexing is scheduled or webhook-triggered on each catalog update.
Q05 Who owns the data, the models and the fine-tunes?
+
Who owns the data, the models and the fine-tunes?
You do. Zero training on your data by default. Logs are kept only for the retention you choose, on the same EU infrastructure. Fine-tunes produced from your data are your property and stay on your tenant. We provide an exit clause: full export of weights, indexes and logs on request.
Q06 Do you offer Batcher.ai with the hosting?
+
Do you offer Batcher.ai with the hosting?
Yes. Batcher.ai is our spreadsheet-style interface to run 10,000+ prompts in parallel across multiple models with workflow orchestration. Core features are included free for our hosting clients. It connects to your private inference cluster or to third-party APIs you already use.
Question not answered?
Contact us →CONTACT, FREE AUDIT
Let's talk about your project.
Free audit, reply within 24 business hours.
Tell us about your stack and situation, a Fast-Mage engineer will call you back. No salesperson, no script. The audit stays yours, even if you don't migrate.