Yahuah Bible Platform.

The story

The Yahuah Bible Platform is a multi-platform Bible study system — a website, an admin panel, a mobile app, and a translation engine. It's used today by real people studying scripture.

The hardest part wasn't the app. It was the translation.

The client needed the Bible translated into French, German, Italian, and Brazilian Portuguese. The catch: every Hebrew or Greek "sacred noun" — the original-language names of God, places, and concepts — had to stay exactly where it appears in every verse. Standard translation APIs would happily move them, drop them, or quietly distort them. For a sacred text, that's unacceptable.

So I built something custom.

What I built

The full system, across three phases:

• A Laravel-powered website with the complete Old and New Testament, inline Strong's numbers, parallel translations, cross-references, and search.
• An admin panel for uploading and managing translations, lexicons, and Strong's entries — with real-time sync to the web app and PWA.
• An offline-first PWA, wrapped in a Flutter shell so it ships as a native app on iOS, Android, Windows, and macOS.
• An AI translation pipeline that runs on GPU servers and translates 60,000+ verses with zero errors.

The pipeline is what I'd call the flagship.

How the translation pipeline works

Plainly: I built a machine that reads the English Bible, finds where every sacred Hebrew/Greek word sits in each verse, lets an AI translate the surrounding language, and then puts the sacred words back into the exact right slots in the new language.

The technical version:

• ETL stage. A Python pipeline extracts all 31,000+ verses and builds a high-precision semantic map of where each sacred noun sits. It uses Unicode regex (Hebrew block ֐-׿) to find them.
• Sequence matching. When translating, I use Ratcliff/Obershelp positional word replacement (difflib) to put each sacred noun back in its correct location in the translated verse.
• The "Leviticus Shift" guardrail. A matcher.ratio() similarity check catches misalignments — if the translated verse drifts below 40% similarity to the original structure, it's flagged for human review.
• A three-layer noise filter caught 7,300+ false positives that would otherwise have slipped through.
• DAG orchestration on RunPod. A custom directed acyclic graph coordinates the GPU work. CPU preprocessing runs in parallel, but I gate GPU inference with a mutex so only one pipeline accesses vLLM at a time — no contention, zero idle GPU.
• Async POST thread pool. Translated verses are pushed back to the Laravel API in the background, and the container terminates itself the moment the work is done — zero wasted compute.
• Human-in-the-loop review. All AI output lands in a staging database first. An admin reviews and triggers publish — nothing reaches production until a human says so.

Stack

AI/inference — vLLM (PagedAttention, KV cache), Gemma 4 31B (full-precision, unquantized), NVIDIA RTX 6000 Blackwell 96GB, RunPod serverless GPU.

Pipeline — Python, custom DAG, Ratcliff/Obershelp matching, async thread pool.

Product — Laravel 10, MySQL, Sanctum auth, PHP 8.1+, Flutter (PWA wrapper), Service Workers, IndexedDB, Web App Manifest.

Outcomes

• 60,000+ verses translated end-to-end.
• 54,409 word replacements verified.
• 0% error rate in production output.
• The architecture generalizes — adding a new Latin-script language is now a config change, not a rewrite.

I also published a long-form engineering case study on the pipeline. The Medium piece covers the DAG, the guardrails, and the cost engineering in detail.