A proprietary translation pipeline, with a human signature on top.
Every Verdacert certified translation runs through a pipeline we built specifically for the demands of USCIS, courts, and universities: vision-based document extraction, per-job model routing across the Claude family, Hijri and Solar Hijri calendar conversion as model tools, per-account translation memory, an AI critic pass, and a credentialed native-speaker reviewer who signs the certification. The AI does the work that's solved. A human does the work that isn't.
AI draft in under 90 seconds · Human-reviewed and certified in 14–48 hours · USCIS-acceptance guaranteed or your money back.
Translation agencies are slow because translators are expensive. AI makes them faster.
A skilled human translator produces about 2,000 words of certified-quality output per day. A frontier model produces a first draft of those 2,000 words in under two minutes — accurate enough that a reviewer's job is to verify and correct, not retype. That single shift is why we deliver in hours what marketplace agencies deliver in days, at a fraction of the cost, with the same signature on the certificate.
Four AI layers, one human signature.
Not one model in a wrapper — a pipeline of specialized stages, each tuned to the part of translation it's best at.
Document understanding
A vision-language model parses the source: detects language and dialect, identifies the document type (birth certificate, transcript, family register, court order, nikah contract…), extracts the visual structure — tables, stamps, seals, signature blocks — and segments text in reading order. This is what lets the final PDF look like the original instead of a wall of text.
Per-job model routing
A router picks the right Claude model per document — Opus 4.7, Sonnet 4.6, or Haiku 4.5 — based on source language, document type, and length. Specialty languages (Pashto, Dari, Tigrinya, Amharic, Kurdish…) and complex or multi-page documents go to Opus for peak reasoning. Short single-page common documents may route to Sonnet to cut latency without compromising accuracy. All inference runs under Anthropic's zero-data-retention agreement.
Domain-aware translation tools
The translator runs with specialized tools no general model has out of the box. Hijri and Solar Hijri (Persian / Jalali) calendar dates are converted via a tool call, not computed in the model's head. Arabic-Indic and Persian digits (٠-٩ / ۰-۹) are normalized to Western digits via a tool call so registry and ID numbers don't drift in transcription. Per-document-type prompt addenda tune behavior for birth certificates, court orders, nikah contracts, academic transcripts, and more.
Translation memory + critic pass
Every document you submit feeds a private translation memory scoped to your account or matter — names, dates, addresses, case numbers, and transliterations stay byte-identical across related filings. Before any draft reaches a human, an AI critic re-reads it against the original document image, looking specifically for omissions, transposed digits, and missed stamps. Reviewer edits flow back as per-language, per-doc-type learned corrections that shape future drafts.
On top of all four AI layers sits a credentialed native-speaker reviewer. They read the source against the draft line-by-line, correct anything that needs correcting, make the judgment calls an AI can't (regional dialect, transliteration choices, ambiguity in the source), and sign the certification statement. No translation leaves our system without this step. Ever.
What AI does, and what only a human can.
The structured parts
- Page count, word count, language detection
- Document type recognition and template matching
- Layout reproduction — tables, stamps, seals, signature blocks
- First-pass translation in under 90 seconds
- Translation memory lookup across your prior filings
- Glossary enforcement (USCIS terminology, court conventions)
- Format-specific checks (date formats, address ordering, name conventions)
- Flagging passages the reviewer should look at first
The judgment calls
- Final accuracy against the source document
- Transliteration of proper nouns (which Roman spelling, when sources conflict)
- Regional dialect and country-specific conventions
- Disambiguating handwriting, faded text, partial seals
- Cross-referencing against your other immigration documents
- The signed certification statement
- Customer communication when the source is ambiguous
Why we're faster, cheaper, and at least as accurate.
| Capability | Verdacert | Typical marketplace agency | DIY (Google / raw LLM) |
|---|---|---|---|
| First draft latency | < 90 seconds | 4–24 hours (translator pickup) | < 60 seconds |
| Layout reproduction | Vision-based extraction | Manual retyping in Word | None |
| Calendar / digit conversion | Tool-call (deterministic) | Hand-converted (error-prone) | None — model hallucinates |
| Translation memory | Private, per account / matter | Per-translator, not shared | None |
| AI critic pass before human review | Yes — every draft | No | No |
| Human reviewer signs the cert | Yes — every document | Yes (the only step) | No |
| USCIS / court acceptance | Guaranteed — refund if rejected | Varies; no guarantee | Not accepted |
| Wall-clock turnaround | 14–48 hours | 2–7 days | Minutes (but unusable) |
| Per-page price (standard) | $19 — or $11.95 with your draft | $25–$40 | $0 (not certified) |
A few specifics, for the people who care.
Model routing
Per-job router picks between Claude Opus 4.7, Sonnet 4.6, and Haiku 4.5. The routing key is (source_language, document_type, page_count). Specialty languages and complex / multi-page documents go to Opus; short single-page common documents may go to Sonnet. Decisions are logged for ongoing recalibration.
Vision-based extraction
Source PDF or image → page rasterization → vision-language model (Claude Opus 4.7) for OCR, structure detection, and field extraction in one pass. Native handling for Arabic, Persian, Urdu, Pashto, Dari, Turkish, Tigrinya, Amharic, Kurdish, and more. Output is structured markdown with detected document type, language, dialect, and a field-by-field map for the translator.
Calendar & digit tools
The translator runs with two model tools: convert_calendar (Hijri → Gregorian, Solar Hijri → Gregorian) and convert_digits (Arabic-Indic ٠-٩ and Persian ۰-۹ → Western 0-9). Date and ID-number transcription errors are the most common cause of certified-translation rejections, and we make them deterministic instead of leaving them to the model's head.
Doc-type prompting
Per-document-type prompt addenda layer specialized instructions on top of the base translator system prompt — what to look for on a birth certificate vs a nikah contract vs a court order vs an academic transcript. Each addendum is a curated set of conventions, common pitfalls, and field-naming rules for that document class.
Translation memory
Every translation produces source ↔ target segments stored in a private memory keyed to your account (or a per-matter sub-scope). Subsequent jobs semantically rank the memory against the new source so near-matches surface even when spelling varies, and names, dates, addresses, and case numbers stay byte-identical across a packet.
AI critic pass
Before any draft reaches a human, an AI critic re-reads it against the original document image — looking specifically for omissions (missed stamps, dropped second pages), transposed digits, and incorrect calendar conversions. Approved drafts go straight to the reviewer queue; revised drafts surface the corrections inline.
Learned-corrections loop
Reviewer edits are captured per (language, document_type) and condensed into corrections that prepend to future translator prompts in that bucket. The translator sees 'last time you translated a Pashto nikah contract, the reviewer changed X to Y — do that here.' Quality of the next draft compounds as volume grows.
Reviewer queue
Completed AI drafts enter a queue scored by specialty match, current reviewer load, and customer SLA. Reviewers see the source, the draft, the translation-memory matches, the AI critic's findings, and the translator's flagged segments side-by-side.
Data isolation
Customer documents are never used to train any model. Claude API requests run under Anthropic's zero-data-retention policy — your content is not retained for model training. Per-account translation memory is logically separated. Reviewer-edit feedback flows back as prompt-level guidance, not as training data leaving your account.
Signed receipts
Every completed translation produces an Ed25519-signed JWS receipt containing the order ID, document hash, reviewer ID, and issue timestamp. Anyone can verify the receipt offline against our public JWKS — useful for downstream agencies that want to confirm a certificate is authentic.
AI speed without AI surveillance.
- Zero-data-retention agreement with our model provider (Anthropic). Source documents and translations are not retained for model training.
- Per-account translation memoryis logically isolated — your account's memory is never queried or surfaced for another customer.
- AES-256 encryption at rest, TLS 1.3 in transit; documents purged from active storage 7 years after completion or on request.
- SOC 2 Type II audit in progress; controls already aligned. HIPAA-aware handling for medical records.
- US-based infrastructure (Vercel + Neon + Cloudflare R2). No data crosses borders for the translation step.
Full details: Privacy policy · Quality & trust · Verifiable certificates.
Questions about the AI layer.
Is the translation done by AI or by a human?
Both — in sequence. The AI pipeline extracts the source, produces a complete first draft, runs an automated critic pass against the original document, and pre-flags ambiguities. A credentialed native-speaker reviewer then reads the source against the draft line-by-line, edits anything that needs editing, and signs the certification statement. The signature on the PDF is a human's; the speed is the AI's.
Which AI models do you use?
We route across Anthropic's Claude family — Opus 4.7, Sonnet 4.6, and Haiku 4.5 — choosing per job based on source language, document type, and length. Specialty languages and complex / multi-page documents go to Opus; short single-page common documents (birth certificates, passports) may route to Sonnet for lower latency without compromising accuracy. Claude is the strongest model for the language pairs we serve and we use it exclusively.
What's actually proprietary about your pipeline?
Several pieces no general-purpose model gives you out of the box: (1) Hijri and Solar Hijri (Persian / Jalali) calendar conversion runs as a model tool, so the translator doesn't compute date conversions in its head. (2) Arabic-Indic and Persian digit normalization is also a tool call — hand-transcription error rate on these is real. (3) Per-document-type prompting tunes the model's behavior for birth certificates vs court orders vs nikah contracts. (4) A per-account translation memory keeps names, dates, and case numbers byte-identical across every document in a matter. (5) An AI critic re-reads the draft against the original document image before it reaches the human reviewer. (6) Reviewer edits feed back into per-language, per-document-type prompt guidance for the next translation.
How is this different from Google Translate or running a model myself?
Three things. (1) The pipeline — extraction, calendar / digit tool calls, doc-type prompting, translation memory, critic pass — is purpose-built for certified translation and is what makes the output USCIS-acceptable. Raw model output is not. (2) Translation memory shared across your matter — names, dates, addresses, case numbers stay identical across every document in the same filing. (3) A credentialed human reviews and signs every output — USCIS, courts, and universities do not accept raw AI translations, and we wouldn't ship one if they did.
Will USCIS or a court accept a translation produced this way?
Yes. The acceptance criterion under 8 CFR § 103.2(b)(3) is the signed certification statement from a competent translator, not the drafting method. Our reviewers are credentialed, read every translation against the source, and sign. The certification is identical to what a manual-only agency produces. Every certified translation we've issued has been accepted by USCIS — and the acceptance guarantee means we re-do any rejection at no charge, no questions.
Do you train models on my documents?
No. Customer documents are never used to train any model. We use Anthropic's Claude API under their zero-data-retention policy — your source content and the generated translation are not retained for model training. Reviewer edits feed back into prompt-level guidance (not model fine-tuning) so future translations of the same document type benefit from past corrections, with no document content leaving your account.
How fast is it actually?
Drafts complete in under 90 seconds for typical single-page documents and under 5 minutes for 20-page packets. The human review step is what determines wall-clock turnaround: 14 hours on Rush, 24 hours on Express, 48 hours on Standard. The AI is not the bottleneck; queue depth at the reviewer pool is.
Can I get the AI draft via API without the human review step?
Two ways. (1) Our Review & Certify tier ($11.95/page) accepts a draft you've already produced, has a reviewer compare it line-by-line against the source, edit, and sign. Same legal acceptance, ~50% the price, ~3× the speed of from-scratch certified translation. (2) For non-certified business translation — marketing copy, internal documents, contracts for review — we'll soon offer an instant AI tier with no human review. That output is clearly labeled 'Machine Translation — not certified' and is not suitable for USCIS, courts, or anywhere a certification is required.
Ready to start? Upload your document for an instant quote.
Standard delivers in 48 hours; Express in 24; Rush in 14. USCIS-accepted, or your money back.