How it works
One funnel: measure first, then fix, then certify. Every step is an artifact your organisation can file, audit and compare.
Step 1
The source document and its translation, in PDF or Word. Lexmetra is built for legal volume: a two-page power of attorney or a court file of hundreds of pages, uploaded the same way.
Large documents are split and analysed in parallel, so a 60,000 word case file does not queue behind a contract. Capacity follows demand automatically; there is no infrastructure on your side.
The focus languages are English, Spanish, French and Italian, and the engine is model-based, not rule-based: it can work with virtually any language. Working in another pair? We want to hear from you.
Step 2
Lexmetra reads the full document pair and returns a Score from 0 to 100. The formula is severity-weighted and structured on industry error taxonomies: critical errors cost more than minor ones, and density matters, so a long clean document is not punished for its length.
The scoring is deterministic: the same findings always produce the same score. That is what makes quality comparable across documents, translators and vendors, and it is why the Score can anchor a service-level conversation.
Calibration is human: the formula is tuned with professional sworn translators, and validated on real court documents with findings verified against source by human audit.
The engine underneath is model-agnostic: Lexmetra runs on the strongest closed and open-weights models (Anthropic's Claude, OpenAI's GPT, Meta's Llama, Mistral and Moonshot's Kimi among them), and adapts to your preference: a provider your organisation already trusts, a commercial agreement you hold, or open weights under your own governance.
The ground truth
There is no better reviewer than the history of your own revisions. Every analysis retrieves from your firm's translation memories, partitioned per client, so findings follow the terminology, style and precedent you have already established with that client: not generic AI opinion.
New clients start from a curated standard memory set, and every delivered revision feeds the next one. Your memory becomes a compounding asset.
Step 3
The complete report: every finding with severity, category, excerpt and suggested fix. Bilingual (English and Spanish) and downloadable as Word.
Two views by design: a high-level summary for the end client, full detail for the reviewing translator.
An expert translator corrects the document, assisted by the report. Orders flow through a shared queue with atomic assignment, delivery tracking and automatic client notification.
An add-on to Human Review for documents that must stand before authorities: the deliverable carries the certifying translator's seal.
Human review is part of every deliverable tier. The AI proposes, the expert owns the result.
Operations
The workflow ships complete: credit-based billing with prices you configure, friendly order IDs your team can quote, monthly consumption counters, and full transparency for administrators. Your ticketing (email or Zendesk) is notified automatically as orders move.