Platform overview
The whole AI lifecycle in one set of records.
EUVYRA covers what teams usually assemble from an experiment tracker, a job scheduler, a serving stack, an LLM observability tool and a governance spreadsheet. The pieces share one data model, so “what runs in production, and why was it allowed to” becomes a query instead of a project.
What EUVYRA does at each stage
Each stage writes records that the next one reads. A release can only be approved against an evaluation of the exact model version, and the gateway only serves releases that were approved, so the evidence for an audit is a by-product of shipping.
| Stage | What you get |
|---|---|
| Plan | An inventory of AI systems, including the ones you buy, with purpose, owners and role under the AI Act; an intake for new use cases with an AI Act pre-screen. Governance |
| Data | Versioned datasets with cards and content hashes, references to files that stay in your own buckets, PII redaction, retention per project and GDPR erasure. |
| Train | Runs, metrics, media, sweeps, reports and model graphs; jobs on your GPUs with queues, quotas and cost; cloud machines within budgets. Training |
| Evaluate | Versioned test suites with server-side evaluators, regression gates for CI, blind pairwise review, annotation queues and online scoring of production traffic. Evaluations |
| Approve and release | Pre-flight checks, sequential approvals bound to people's functions, direct, canary and shadow rollouts, promotion and rollback. |
| Serve | Endpoints with vLLM, llama.cpp, any container or EU providers behind an OpenAI-compatible gateway with redaction, tracing, budgets and rate limits. Serving |
| Operate | Node and GPU inventory with utilisation history, maintenance mode, excluded nodes, and pools of cloud machines sized to the queue. |
| Monitor | Traces of every call, monitors for errors, latency, cost, quality scores and drift, and incidents with reporting clocks. |
| Govern | AI Act risk classification, controls checked against evidence, signed evidence packages, and assurance across the AI Act, NIS2, DORA and the GDPR. |
| Secure | Roles, governance functions, scoped API keys, single sign-on and SCIM provisioning, a hash-chained audit log and an outbound allowlist. Security |
| Work with assistants | An MCP server, so AI assistants can work in EUVYRA as the person who connected them, while approvals stay with people. Developers |
Two anchors, one chain of evidence.
Everything hangs off two records. A project owns runs, traces, datasets, test suites, evaluations, models, prompts and monitors. An AI system is the unit a regulator cares about: it links one or more projects and owns risk assessments, controls, releases, incidents and evidence exports.
Lineage is a chain of references that EUVYRA walks on demand:
- a release points to a model version or prompt version;
- the model version points to the run that trained it, its artifact digest and model card;
- the run points to its git commit, configuration and the dataset versions it read, each with a content hash;
- the release lists the evaluations it passed and every approval step, with who signed and when.
That is how EUVYRA answers the questions auditors and incident responders ask: which production models were trained on dataset X, did their last evaluation pass, and what changed since they were approved.
What you install
A small number of parts that your platform team already knows how to run. The security page covers how they are locked down.
API server
One FastAPI application for ingestion, governance, evaluators and evidence signing, with migrations that run on start or as a Helm hook.
Worker
The same image running background jobs: retention, monitors, online evaluation, webhooks, cloud machine orders. Several replicas never run a job twice.
Console
The web app people use. It talks only to its own origin and passes API calls through, so the browser needs no CORS and the session cookie stays first-party.
PostgreSQL
All records, including traces, metric points and the audit chain. A database trigger rejects updates and deletes on the audit table.
Object storage
Any S3-compatible store for artifacts and evidence packages, with server-side encryption. The Compose stack ships one.
Agent
Runs next to your GPUs in a Kubernetes namespace or on a Docker host, pulls jobs and endpoints, and reports state, logs and GPU readings.
Neutral by design.
EUVYRA has no model of its own to sell and no cloud to fill. You choose where training runs, which models you serve and which providers you call, and you can change each choice without changing the platform.
- Open models on your own GPUs with vLLM or llama.cpp, or any container that speaks HTTP.
- EU model providers behind the same gateway, reachable only when their host is on your allowlist.
- Machines from Hetzner, Scaleway, OVHcloud, AWS or Google Cloud, with EU regions offered by default.
- OpenAI-compatible APIs and OpenTelemetry on the way in, so your code is not tied to EUVYRA either.
from openai import OpenAI
client = OpenAI(
base_url="https://euvyra.example.eu/v1",
api_key="euv_...",
)
reply = client.chat.completions.create(
model="radiology-triage", # an AI system
messages=[{"role": "user", "content": "Hallo"}],
)Good to know before you rely on it
- EUVYRA collects evidence and supports review. It does not certify conformity with the AI Act, NIS2, DORA, the GDPR or ISO/IEC 42001, and its risk classification and regime scoping are documentation aids, not legal advice.
- Reports to authorities are drafted and tracked with their deadlines, not sent. Authorities have their own portals.
- Traces, spans and metrics live in PostgreSQL. That is comfortable into the tens of millions of spans.