Closed testing · invite only

Every item aligned. Every number verified.

Transcripts, certificates and notarial records — translated with the table structure intact, every digit conserved and machine-checked, every term rendered the same way throughout, and the source and translation matched item by item. Anything unaccounted for stops the job for human review; nothing passes silently.

PDF / PNG / JPG · zh / en / es / fr / de / ja, any direction · artifacts deleted after 30 days

A 47-row Chinese academic transcript beside its English translation, tables aligned
Academic transcript, 47 courses in two column groups · Chinese → English · synthetic sample

Item by item, the translation answers to the source

English course names run two to three times longer than the Chinese, yet every row stays a row and every column stays a column — the reader finds any item on both sides by its place in the table. Every screenshot here is real pipeline output; the documents are synthetic fixtures — a real customer document never becomes marketing material.

Course rows in Chinese and English side by side, each row aligned with its translation
One semester of course rows — each item in its place on both sides

Numbers are verified by code, not trusted to a model

Before translation, code freezes every critical value — grades, credits, IDs, dates, amounts — into immutable tokens the model cannot touch; afterwards code verifies each one back into place, and a conservation gate compares the full set of digits on both sides. The model never gets a chance to re-type your GPA.

The transcript's summary numbers — total credits, GPA, ranking — identical in source and translation
Totals, GPA and ranking — identical on both sides, checked digit by digit

Seals are described, never copied

Certified translation has one hard rule: stamps, embossed seals, seam seals and signatures are described in brackets and never pasted in as images — copying the red seal across is a leading reason documents get rejected. CertTwin reads the text inside the seal, translates it, writes a proper italic bracketed note, and makes you approve each one.

A red round seal on the source, and the bracketed description in the translation
The source's red seal → the translation's bracketed description

Click anything to edit it — both sides light up

Click any text on the page and the matching place highlights on the source and the translation at once, so you never hunt for correspondences. Edits redraw immediately, seal descriptions need your explicit approval, and every action lands in an audit log.

The same content highlighted on both the source and the translated page
One selection, highlighted on both pages

Every character is accounted for

Zero silent loss is a checkable property, not a slogan: extraction is reconciled against the document's own text layer, anything the model missed is recovered from it, repeated terms must translate identically everywhere, and every draft runs all nine checks. If any one fails — including two models reading the same scanned digit differently — the job stops for human review instead of going out.

CoverageEvery page extracted, native text layer fully reconciledpass
Visual elementsSeal and signature descriptions await your approvalreview
Number conservationThe set of digits matches between source and translationpass
Block completenessNo semantic block dropped or blockedpass
Language purityNo source-language residue or Chinese punctuation left behindpass
Term consistencyA repeated source string translates identically everywherepass
Minimum font sizeNothing squeezed below the readable floorpass
Layout collisionsGlyph-level scan; any overlapping text is heldpass
Side-by-side reviewHuman sign-off — never auto-approvedreview

From one document to a batch

Translation agencies, law firms and admissions services process documents by the folder, not one at a time. The pipeline behind CertTwin is queue-based and runs pages in parallel — and every document in a batch passes the same nine gates individually, so scale never dilutes scrutiny. Batch intake is in invited testing: if you process volumes, write to us and we will set you up.

Three steps, and you have the last word

1

Upload

A PDF, a scan or a phone photo. Files go straight to private storage, with no third-party middleman.

2

Translate with protected fields

IDs, dates and amounts are frozen before translation and verified back afterwards. On scans, a second model re-reads the digits.

3

Review side by side, then export

Check each place, edit in line, approve the seal descriptions, and export the review draft as a PREVIEW_ONLY PDF.

Documents people bring us

Tuned for the formatted paperwork behind study abroad, immigration, visas and notarization.

Transcripts / diplomas Birth / marriage certificates Notarial certificates Bank statements Police clearance Household register / ID Business licences Patents and awards

Questions people ask

Is this a certified translation?

No. Every page is stamped PREVIEW_ONLY — a review draft meant for human verification and whatever certification step follows. It is not itself a certified translation and does not guarantee acceptance by any authority. Whether a translator's declaration gets attached is your call.

What happens to my documents?

They are used only to produce your translation, processed by Google Cloud (Vertex AI) in the United States, and the artifacts are deleted after 30 days. This site sells nothing and never asks for passwords, card numbers or ID numbers.

Which languages?

Chinese ⇄ English today — the pair behind most study-abroad and immigration paperwork. More are on the roadmap.

Why invite-only?

The product is in closed testing, and we want every problem a real document exposes to actually get fixed. Email us with your document types and purpose to join.

Want to try it on your own documents?

Free during closed testing. Send us an email describing what you need translated and we will open an account.