Tiv Heritage Archive

Designing a translation engine for a low-resource African language like Tiv required an entirely custom pipeline rather than relying on generic NLP tools. The solution was an 8-layer architecture that queries exact proverbs first (preserving cultural nuance), then curated phrase pairs, partial matches, dictionary words, category lookups, word-by-word grammar reconstruction, and finally the NLLB AI model as a last resort. This layered approach ensures that culturally significant phrases always take precedence over statistical AI guesses, while a missing-words queue and community suggestion workflow continuously expand the dictionary without admin bottlenecks.

Building a fully custom MVC framework from scratch — including a regex-based router, CSRF middleware, file-based caching, and an ORM-like base model with full-text search — meant every security concern had to be handled manually. Session ID regeneration every 5 minutes, bcrypt password hashing, progressive IP blocking, Content Security Policy headers, and JSON delta audit logs were all implemented without a framework scaffolding these defaults, requiring careful design to avoid common vulnerabilities.

Modelling the knowledge graph across 25+ tables with a generic knowledge_links pivot presented a schema challenge: any content type (plant, festival, food, proverb) needed to be linkable to any other, without a proliferation of join tables. The solution was a polymorphic source_table/source_id + target_table/target_id design, enabling relationship types like "plant_used_in_festival" or "ingredient_in_food" across all categories through a single table.

Supporting the freemium translation model required tracking daily usage for both guests (IP-based) and authenticated users (DB-based) while simultaneously recording payment subscriptions, credit balances, and expiry dates. The translation_logs, translation_payments, and site_settings tables were designed together so that the engine checks payment status first, falls back to the daily limit, and logs every request — giving the admin full analytics on which words are missing and where the translation pipeline fails most often.