An Arabic-language education initiative
A corpus of 676 Arabic doubled-verb conjugations needed to be explorable by researchers and students alike. Existing academic tools required CSV wrangling and produced static PNGs that nobody could interact with.
What they came to us with.
The initiative maintains a 676-entry corpus of Arabic doubled-verb (fi‘l muda‘‘af) conjugations indexed by fāʾ / ʿayn letter combinations. Existing academic tooling required manual CSV wrangling and produced static PNG outputs — fine for a paper, useless for a student trying to explore the space.
They needed something researchers could publish from and students could browse — same data, two audiences — without forcing either to write Python.
How we built it.
The front end is an interactive Next.js matrix of all 676 entries. Click any cell and the platform surfaces its conjugation variants, phonetic attributes, and the cluster it belongs to. Behind it sits a Python pipeline that extracts phonetic features from the sifat strings, computes UPGMA hierarchical clusters using Jaccard distance, reduces to 2D via UMAP for visualization, and selects k by silhouette score.
The deliberate choice here was to auto-discover phonetic features rather than hand-code them. The initial hand-engineered feature set was concealing real cluster structure; letting the data reveal what was load-bearing surfaced fāʾ / ʿayn co-occurrence patterns that the hand-coded version had flattened out.
What shipped. What changed.
Time to explore corpus
Feature discovery method
Cluster count selection
Output reproducibility
Keep reading.
Education — A Gulf-region research university digital-humanities centre
Bilingual Knowledge-Graph Platform for 200 Years of Gulf Archival History
Thirty-three PDF volumes of 1622–1810 Gulf history — 1.94M words over 5,657 pages — existed only as unsearchable prose, alongside a parallel Arabic edition whose text layer was unrecoverable. No structure, no entities, no map, no way to cite a passage.
- 99.99% source-text capture across 33 volumes — 1,939,480 of 1,939,670 words, every drop classified as page furniture
- A knowledge graph of 2,687 entities, 16,377 events and 19,814 typed relations drawn from 4,328 records spanning 1622–1810
- Arabic edition taken from 524 to 3,985 records at 1:1 parity with English, with only 1.9% machine-translated
Other — A publishing-adjacent services company handling 500+ DOCX assets per month
Editorial Workflow Platform for Multi-Volume Document Production
Editorial ops was running 11 separate tools to move a document from intake to edit to merge to format to export. Turnaround was unpredictable and any one tool outage stalled the whole pipeline.
- 11 editorial tools consolidated into 1 workspace
- Heavy document merges moved off the request path (no more timeouts)
- Median turnaround per document: 3.5 days to 10 hours
Want the same outcome for your team?
Tell us where you are now. You'll get a fixed price in writing before any work starts.