Sovereign AI for Greek business, priced on outcomes, never metered per token.
Data and compute under Greek and European control.
Open-weight models, fine-tuned on real Greek workloads and served from GPUs in Greek data centres. Voice agents for contact centres, an AI receptionist for small businesses, AI Employees for the mid-market, and the inference core underneath.
The platform at a glance: pick a layer or follow a call.
GPUs
On a callThe whole call runs on GPUs in Greek data centres or on your own premises.
Inference
On a callClio transcribes the caller, Metis chooses the tool and Calliope speaks the reply.
Router
On a callEach turn is routed to the model that fits the task, based on confidence.
Company Brain
On a callThe agent reads what it needs from the Brain: who is calling, what they have and what is open.
Governance
On a callThe rule gate checks the proposed action against your rules and logs the decision.
Agent harness
On a callThe agent runs a skill, such as looking up an order: a typed, tested unit, not a free-form prompt.
Apps & channels
On a callA customer calls over SIP. The voice agent app owns the conversation and decides when to hand off to a person.
Our inference stack runs in production on contact-centre workloads for a Greek telecom operator and a Greek bank; the AI Employee platform is live at a leading Greek publisher. See the measurements
One sovereign core, four products
The products share one Company Brain, meaning your entities, your rules and an audit trail of every decision. You buy an outcome: an answered call, a booked appointment, an answer that matches your books. Never metered per token.
Seven layers, from the GPU up
A call, a chatbot, a coding session: every request crosses the same seven layers. Bring your own brain, model, skills or apps. Governance stays in place.
- Layer 1 / 7 · Sovereign Inference Core
GPUs
Dedicated GPU capacity in Athens data centres operated by Greek providers. Capacity is sized to the workload, so you know what one GPU carries before you commit.
- Hardware
- NVIDIA H200
- Location
- Athens data centres, Greek providers
- Capacity
- Dedicated, or on your premises
200 / GPUconcurrent Greek call summaries within a 2-second KPI
423 / GPUconcurrent Greek text-to-speech streams
Business outcomeData and compute stay under Greek control, at a capacity cost you can plan.
Bring your ownRun the same stack on your own premises.
- Layer 2 / 7 · Sovereign Inference Core
Inference
Four models of our own, each an open-weight base with a fine-tuned adapter, served on vLLM. The model is sized to the task, which is where the cost difference against a frontier API comes from.
- Serving
- vLLM · FP8 · prefix caching
- Instant LLM
- Iris · 8B
- Thinking LLM
- Metis · 12B · 26B
- Speech to text
- Clio · 1.5B
- Voice
- Calliope · 3B
- Built on
- Krikri · Gemma 4 · Whisper · Orpheus
≈ 10–250×cheaper than frontier APIs, depending on the task and the frontier model compared. About 30× for Greek call summaries. Estimate.
−45%latency at 200 concurrent requests, from prompt reordering alone
Model cards · How we measured this
Business outcomeFrontier-level answers on Greek documents (4.41 / 5) from Metis 12B, a model you can host yourself.
Bring your ownServe your own fine-tuned model next to ours.
- Layer 3 / 7 · Sovereign Inference Core
Router
One endpoint in front of every model. A request starts on a small model and moves to a larger one only when confidence is low.
- API
- OpenAI-compatible
- Routing
- By confidence, small model first
- Alternatives
- Llama · Qwen
- Feedback
- Failed traces become skill fixes or fine-tuning examples
Business outcomeThe large model is used only when it is needed, and you change models without touching your application.
- Layer 4 / 7 · Business context
Company Brain
What the agents know about your business: customers, products, contracts and calendars as typed entities and relations, kept current through connectors to your systems.
- Data model
- Typed entities and relations
- Store
- Bitemporal: what was true, and when it was recorded
- Sources
- Connectors to your ERP and business systems
- Start from
- Ontology templates
Business outcomeAnswers that match your books, because the agent works from your records and not from what a model remembers.
Bring your ownBring your own brain: your knowledge graph or data model plugs in here.
- Layer 5 / 7
Governance
The control layer between a model and anything it does. The model proposes; the rules decide.
- Rule gate
- Checks every proposed action before it runs
- Audit trail
- Append-only, hash-linked record of every decision
- Compliance
- AI Act dossier and model card for every model
Business outcomeEvery automated decision can be shown to a customer, an auditor or a regulator.
Always onStays in place whatever brain, model or app you plug in.
- Layer 6 / 7 · Skills & Agents registry
Agent harness
Where a model becomes an agent. A small model with tested skills does work that would otherwise need a much larger one.
- Agent
- Role + skills + policies
- Skill
- Typed contract, deterministic code, tests
- Release
- Signed skill packs
- Evaluation
- Paired: with and without the skill
131 / 190booking and service tasks completed correctly by Metis 26B
Business outcomeBehaviour you can review before it ships: every role is composed from tested skills.
Bring your ownAdd your own skills to the registry.
- Layer 7 / 7 · Factory
Apps & channels
Where the work arrives. The Factory turns a request or a “hire” into an App Manifest and provisions it; publishing is provisioning.
- Workloads
- Calls, chatbots, coding sessions, back-office work
- Channels
- Voice, WhatsApp, web, telephony (SIP)
- Build
- App Manifest through the Factory
- Pricing
- Per outcome, never per token
Business outcomeYou buy an outcome: an answered call, a booked appointment, an answer that matches your books.
Bring your ownBuild your own apps on the OpenAI-compatible API.
Measured, not claimed
We publish what we measure on our own GPUs, with the stack we ship.
We publish what we measure, and we study how AI is measured
Eleven papers, submitted for peer review, on training small models, measuring them honestly and learning from production data safely.
Explore the researchYour data stays in Greece. By rule, not by promise.
Customer data and all serving stay in Greek data centres. US APIs are never in the serving path. Frontier models are used only as judges and teachers in evaluation.
The rule applies to every product served from the Hellas Inference core. AI Employee runs today on EU cloud with third-party model APIs and moves onto the Greek core in 2027; the rule applies to it from that migration.
Providers
What runs where today: Contact-centre workloads on dedicated capacity in Athens data centres. AI Employee on EU cloud until its 2027 migration to our Greek core. Training and evaluation at Greek sites and through PHAROS AI-Factory access; benchmarks on rented GPUs use non-customer data only.
- Dedicated GPU capacity in Athens
Dedicated GPU capacity in Athens data centres operated by Greek providers.
Production serving, contracted per customer - PHAROS AI Factory / DAEDALUS (GRNET, EuroHPC)
Greece's national AI factory at Lavrion, for fine-tuning and large evaluation runs through AI-Factory access.
Training and evaluation · subject to access grant - Owned H200 nodes
Owned H200 nodes in Athens colocation as capacity insurance.
Planned 2027
Call our models from your own code
An OpenAI-compatible API on GPUs in Greek data centres. Price your workload next to frontier APIs, then create an early-access account. An administrator reviews every account and grants access.
We run Greek business workloads in production
Hellas Inference runs Greek business workloads in Greek data centres. We take the best open-weight models, fine-tune them on Greek speech and text, measure them under production conditions and sell what they deliver. Based in Athens.
CTO / ML lead and Greek-speaking ML engineers
If you have shipped speech or language models to production and want to do it for Greece, write to us.
Start with one workload.
Tell us the calls, documents or questions you want handled. We measure it on your data before you commit.