OUP Translation Platform

Secure, AI-assisted publishing for South African languages — end-to-end workflow from authorised upload to linguist-approved export.

Content library Steering committee demo
End-to-end translation workflow (conceptual)

Import authorised content

Upload OUP material individually or as a ZIP (compressed archive) batch. Tag subject, CAPS (Curriculum and Assessment Policy Statement) grade band, document type, and target language before content enters the private library.

Drop PDF, Word, ZIP, or folders here

Drag from your desktop — or click anywhere in this area to browse

.pdf .docx .doc .zip

Detected from upload filename — overwrites manual entry when a file is selected

Detected from upload filename — overwrites manual selection when a file is selected

Matches content types defined in the OUP translation platform proposal.

Eight official target languages — English source; phased rollout recommended in proposal.

Encrypted OUP tenant Egress locked — approved AI endpoint only No public internet as corpus source Provider no-training / retention controls

Ingest & build the private index

Content is validated, extracted, chunked, and indexed for meaning-based retrieval — entirely within the OUP tenant. Progress on the left; live activity log on the right.

Preparing ingestion… 0%
0
Pages processed
0
Segments extracted
0
Index chunks
0
Vector embeddings

Processing pipeline

Live activity log Streaming

Simulated AI draft speed

224
Pages per hour
3.73
Pages per minute
4
Parallel threads
80 min
300-page bulk est.

Translation worker threads

4 workers × 56 pp/hr = 224 pp/hr · 4 spare threads available for bulk scaling

Bulk estimate (simulated): 500 pages ≈ 134 min · 1,000 pages ≈ 268 min · 4 parallel translation threads active

Live throughput for English → target language pair (machine draft stage only)

AI-assisted drafting (simulated)

Translation memory and glossary first — then context pack assembly from OUP-only sources — then enterprise AI draft via server-side API.

Initialising translation job… 0%
0
Segments drafted
0
TM (Translation Memory) matches
0
Context excerpts
0
QA (Quality Assurance) checks run

Drafting pipeline

Draft engine log Idle

Simulated machine draft speed

224
Pages per hour
3.73
Pages per minute
4
Draft threads

224 pp/hr · 3.73 ppm · machine draft only (per proposal) — linguist review is additional

Language pairEnglish → Afrikaans
AI routeServer-side API (Application Programming Interface) · walled garden
Workers4 parallel draft workers

Context pack — retrieved from OUP library (RAG — Retrieval-Augmented Generation)

Reference excerpts from the private library used to guide the AI draft — retrieved for meaning, not validated as final output.

Automated QA (Quality Assurance) checks (per proposal)

Risks surfaced to reviewers before human sign-off — never silently corrected.

Draft output — segment 1 of 24

Source: The plant needs sunlight to grow.

Linguist review workbench

Mandatory human sign-off with MQM (Multidimensional Quality Metrics)–style error rubric (per proposal). Automated QA (Quality Assurance) runs first; linguists score each segment before approval. Approved pairs commit to translation memory with full audit trail.

Automated QA (Quality Assurance) — pre-review results

English (source)
Afrikaans (draft)

MQM (Multidimensional Quality Metrics)–style error rubric (per proposal)

Confirm each category before approval. Flagged issues route to terminology control and pilot QA (Quality Assurance) metrics — indexed for retrieval within the OUP tenant only.

4/5

Quality & control loop

How automated QA (Quality Assurance) and human MQM (Multidimensional Quality Metrics) scores feed TM (Translation Memory), glossary, audit, and pilot acceptance — within the walled garden.

No segments approved yet — quality signals will appear here.

Export approved package

Approved segments are written to translation memory and exported in formats agreed with OUP publishing — with full audit trail.

0 segments approved · ready for export to Afrikaans