- Outside — never used
- Public AI APIs
- Third-party model providers
Ask the question.
Get the finished work.
Not a link to a file — the deck, the table or the answer, ready to send, with the source behind every figure. DOYNA reads your company's email, SAP, contracts and files. And none of it reaches a public AI.
Inside the perimeter
- Sources: email, SAP, contracts, files
- Vector index
- Model inference
- The answer, with citations
- Answers shaped from real runs on a customer corpus, from a test account. Figures are round on purpose — they are examples, not customer data.
What people ask on a Monday morning.
Open one. What sells is not the number — it is the shape of the answer: how many sources it read, what it chose and why, and where you can check.
+−What was our EBITDA last year?
It read
Dozens of sources, several tools. It found four different EBITDA figures, in different documents.
It returned
Not one at random: all four in a table — say 12, 14, 15 and 18 million — with the file and the exact cell for each, then which one is right and why.
Why it sells
A person would have quietly picked one. This shows you there were four — the thing that would have cost you in the meeting.
+−How much are we owed, and how much is overdue?
It read
The receivables report for the month, across several sources.
It returned
1,000,000 total, of which 300,000 over 30 days and 700,000 inside the window — every figure carries a citation mark you can check one by one.
Why it sells
A direct answer to a question you ask a person today and get back in two days.
+−What are our largest costs?
It read
The expense accounts, ordered by total turnover, for the whole year.
It returned
Top five, with a chart and a sortable table. Raw materials first — say 400 million in total turnover.
You got
↓ export
+−How much of our profit comes from one plant?
It read
Two separate documents. The answer was in neither of them alone.
It returned
20% — dividing a figure in one document by a figure in another, with both sources shown.
Why it sells
This is not search. It is the operation a person would have done with two files open and a calculator.
+−Add up the turnover of the companies in the group.
It read
The statements of each entity.
It returned
80 million RON — and it checked for itself that this matches the consolidated figure before handing it over.
Why it sells
It does not leave you to discover the mismatch.
- The difference between a search engine and a colleague: one gives you the link, the other gives you the work.
You leave with the file, not a link.
What you actually get at the end of a conversation.
Decks with figures tied to their source, not screenshots.
The same material, fixed for sending on.
Direct, with a link to the document behind every claim.
- The person who signs the contract is not the person who uses the product. Each needs their own question.
One install, four jobs.
Finance director
„Where is the money leaking that nobody sees?”
Legal
„What did we sign, and what expires without anyone asking us?”
Operations
„Where is the order, and who is it waiting on?”
Chief executive
„What do I need to know before Thursday's meeting?”
- We do not name the customer on this site.
A paying customer, live in production.
Not a pilot sandbox: search, the knowledge map and the agent that does the work are in daily use inside a Romanian company, on their own data.
In production
A Romanian enterprise
890+ people, one private AI over their email, contracts and files. Ask us on a call and we will walk you through what it does for them.
- A published figure with a named source. We would rather quote the study that says most of these fail than pretend the risk is not there.
95% of enterprise AI pilots fail. Doyna is built to be the 5%.
MIT NANDA’s report The GenAI Divide: State of AI in Business (2025) found 95% of enterprise GenAI pilots deliver no measurable P&L impact — almost always from integration and trust, not model quality.
Why Doyna lasts
- Runs on your data — answers cite source, page and date
- Abstains when unsure, instead of guessing wrong
- Does the work inside the flow — so it earns daily use
Why 95% of pilots die
- Generic models, never grounded in your data
- Confident wrong answers, with no source to trust
- Bolted on beside the work — adoption quietly dies
Pilots die on trust. Doyna shows its sources — or says it doesn’t know.
- Four demos on synthetic data. Everything stays in your browser — nothing is sent anywhere from this page.
Try it here, before you ask anyone for a demo.
The real DOYNA runs on your own email, documents, contracts and SAP. Pick a tab.
- No competitor is named. The argument holds; comparative advertising against a named brand does not need to be part of it.
Where the difference sits.
A table, not cards. Every row has to be defensible before it is published.
| Criterion | DOYNA | Cloud assistant |
|---|---|---|
| Where inference runs | DOYNAOn a dedicated server, one company only | Cloud assistantThe vendor's cloud |
| How you buy it | DOYNABox once, software yearly | Cloud assistantPer seat, forever |
| Romanian, natively | DOYNAYes | Cloud assistantPartly |
| A citation on every answer | DOYNAYes | Cloud assistantVaries |
| Trains on your data | DOYNANever | Cloud assistantOpt-out, usually |
| Who else shares your instance | DOYNANobody — single-tenant | Cloud assistantEvery other customer |
| Email, documents and scanned archives | DOYNAOne index | Cloud assistantSeparate tools |
| A map of who connects to what | DOYNAGraph index, fused with two others | Cloud assistantNot available |
| Agents that finish the task | DOYNAYes | Cloud assistantSuggestions |
| Runs with no internet at all | DOYNAAir-gapped | Cloud assistantNo |
- What you pay for the server: the supplier's price plus €5K or 10% handling, whichever is higher, ex-VAT. Supplier prices from October 2026. GPU prices are moving fast, so every server is re-quoted at order.
Own the server. Don’t rent someone else’s.
Renting AI means paying for somebody else's infrastructure, permanently.
| Tier | Configuration | Price | Status |
|---|---|---|---|
| Starter | Configuration1× RTX PRO 6000 workstation | Price~€26K | Estimatedfor up to 10 people* |
| Growth | Configuration1× H200 NVL server | Price~€61K | Estimatedfor up to 30 people* |
| Scale | Configuration2× H200 NVL server | Price~€99K | Estimatedfor up to 60 people* |
| Enterprise | Configuration4× H200 to 8× B300 | PriceOn request | On request60+ people*, up to a GPU cluster; NVIDIA end-user checks apply |
Hardware only, excluding VAT and the DOYNA software subscription. *At today's average use; heavy users, fewer. Sizing is indicative until load-tested.
- This is the one place on the site that does not ask you to take our word for it. Straight from our own analytics.
Real traffic, not screenshots.
Our own site, measured. These figures are read out of our analytics every time the site is built.
Measured 10 October 2026 – 9 November 2026, from our own analytics.
The questions everyone asks before the first demo call.
+−Is this real? Or is everything on this site a mockup?
The traffic figures in the Real numbers plate are ours, read out of our own analytics at build time. The interactive demos on this page are stylized React components built from the real Doyna UI (we can’t show real customer data on a marketing site). Book a demo and we’ll show you the production app.
+−What do “knowledge graph”, “RAG” and “air-gapped” actually mean?
Plain-English glossary. Knowledge graph: a live map of who’s connected to what — which people, deals, contracts and projects relate to each other. In Doyna it is one of three indexes, not the whole system: a vector index holds meaning, a Neo4j graph holds the links, and PostgreSQL holds the values that must be exact, and retrieval fuses all three. Full-corpus search (RAG): the AI reads across everything your company knows before answering, instead of guessing. Entity resolution: figuring out that “Jane Doe”, “J. Doe” and “the CFO” are the same person. Semantic search: search by meaning, not exact words. Single-tenant: your own private copy, never mixed with other customers’ data. Air-gapped: runs with no internet connection at all, fully inside your building. In-perimeter / on-prem: runs on a dedicated server that serves only your company, at your site or in a managed data centre, so your data never reaches another customer or a public AI. Bring your own LLM: plug in whichever AI model you’re allowed to use.
+−What happens to my data if you go out of business?
Nothing. Doyna runs on a server you own, at your site or in a managed data centre — the corpus, the indexes and the configuration sit on its disks and stay there whatever happens to us. Source-code escrow is available on request as part of an enterprise agreement.
+−Do you train on my data? Even a little bit?
No. Not for model improvement, not for cross-customer learning, not for vendor evals. Your embeddings live in your tenant’s Qdrant collection, your prompts and responses are never aggregated across customers. This is contractually guaranteed and technically enforced by single-tenant isolation.
+−Can I try it without talking to sales?
Not today — there is no self-service signup. Every account starts with a short call, because we connect your first source with you instead of handing you an empty product. From there you evaluate Doyna on your own data, on a machine we own at our office, before anything is installed on yours.
+−Do you support Romanian?
Yes, natively — not via translation. The AI assistant, email rules and the entire UI are built bilingually. You can write rules like “nu mai vreau emailuri de la X” and they work. The website has an EN/RO toggle in the top right.
+−What if I want a feature that isn’t in your list?
Tell us. The roadmap is driven by our first customer — feature requests from paying customers get prioritized over everything else.
+−Is support included?
Email support on every plan. Larger teams get priority response (1 business day SLA), a dedicated account manager, a named Slack/Teams channel, and on-call during rollout — ask sales what fits your plan.
See it running on infrastructure like yours.
A technical walkthrough: architecture, deployment model, what a pilot takes.