Leading AI models — run in EU regions, every provider under a data-processing agreement (Art. 28), with EU Standard Contractual Clauses for US providers. Two models independently check form and substance, an arbiter synthesises — transparent about which provider processes what.
Four steps, each logged. Processing in EU regions, every provider under a data-processing agreement.
PDF with a real text layer. Word count is checked server-side against the tier.
Mistral (France · EU) prepares the document into clean, structured text via OCR and Document AI — better legibility and clear structure for the review that follows.
Two models check form and substance, a deterministic engine checks spelling — all write into the same JSON schema. One arbiter synthesises, deduplicates and calibrates severity.
Inline findings, severity in three levels, rationale per passage. Receipt with model versions.
Word count determines the tier — and the review standard scales with the level (Bachelor's, Master's, dissertation). One Stripe payment, one delivery.
Mistral (France · EU) prepares the document into clean, structured text via OCR. Then two models review form and substance — identical JSON schema, low temperature. Spelling and grammar are additionally checked by LanguageTool locally, deterministically.
Bachelor's, Master's and dissertations are not treated the same: the review standard and the wording of the feedback scale with the level — from solid fundamentals and easy-to-apply suggestions to whole-work review (coherence across many chapters, consistency across a large reference apparatus, a publication-ready register).
Document preparation via OCR and Document AI into clean, structured text — better legibility and clear structure for the review models.
Form: citation, formatting, source references, tense.
Second review pass on HIGH (structure, coherence, style, readability); then synthesis on MAX — deduplicates, separates finding from suggestion, calibrates severity.
Maximum model power, processed in EU regions — every provider under a data-processing agreement (DPA). Transparent about where data sits and where the residual risk is.
Lektoria processes with European or EU-region providers. There is a data-processing agreement under Art. 28 GDPR with each one; for the US groups Microsoft and AWS, EU Standard Contractual Clauses and certification under the EU-US Data Privacy Framework apply in addition. Microsoft Azure and AWS Bedrock do not train on inputs; AWS Bedrock does not retain them, Microsoft Azure may store them for up to 30 days for abuse monitoring within the EU. For Mistral, the training exclusion and Zero Data Retention are still to be confirmed contractually.
| Processor | Purpose | Region | DPA |
|---|---|---|---|
| HetznerBackend & storage | FastAPI · Postgres · S3-EU · LanguageTool (spelling, local) | Falkenstein · DE | Art. 28 GDPR ↗ |
| Mistral AIDocument preparation (OCR / Document AI) | Document to structured text (OCR) · European company · ZDR to be confirmed contractually | France · EU | Mistral DPA (Art. 28) ↗ |
| Microsoft AzureReview — GPT 5.4 (HIGH) | Azure OpenAI · EU Data Boundary · no training · storage ≤ 30 days (abuse monitoring, EU) | EU Data Boundary · Sweden | DPA (Art. 28) + SCC + DPF ↗ |
| AWSReview + synthesis — Opus 4.8 via Bedrock (HIGH → MAX) | Amazon Bedrock · no training · no persistent storage (ZDR) | Stockholm / Ireland | DPA (Art. 28) + SCC + DPF ↗ |
| IONOSTransactional email (receipt, delivery) | SMTP — notification + download link | Germany | Art. 28 GDPR ↗ |
| StripePayment | Checkout · Refunds | Dublin · IE | Stripe DPA ↗ |
Processing in EU regions does not in every case fully exclude access by an authority from a third country (e.g. under the US CLOUD Act). We disclose this residual risk and secure transfers via EU Standard Contractual Clauses and the EU-US Data Privacy Framework — details in the privacy policy.
More questions: info@lektoria.eu
model_failure, timeout or budget_exceeded a refund is issued automatically via Stripe.