Exnav29/every-tongue

AI assistant for drafting and checking Scripture study materials in low-resource languages. AMD Developer Hackathon: ACT II. # Every Tongue **Scripture passage in → grounded study material out → English back-translation → reviewer handout — powered by Gemma on AMD GPU.** Every Tongue drafts **reviewable study material from existing human Bible translations** for low-resource languages. **It does not translate Scripture** and **requires native-speaker review** before any use. - 🟢 **Live demo:** huggingface.co - 📦 **Repo:** github.com - 🔬 **AMD/Gemma evidence:** `evidence/amd-gpu/` - 🖥️ **Slide deck:** `docs/Every-Tongue-deck.pdf` Built for the **AMD Developer Hackathon: ACT II** — Track 3 (Unicorn). --- ## ⏱️ 60-Second Judge Demo 1. **Open the live demo:** huggingface.co 2. **Target translation:** leave **Swahili — Kiswahili Neno 2015** (the default). 3. **Passage reference:** type **`John 3:16-17`**. 4. Click **📖 Fetch passage** — the passage loads in English *and* Swahili, pulled from ScriptureFlow (real expert human translations, side by side). 5. Click **✍️ Draft study material** — a grounded study guide is drafted in Swahili and English, with phrases flagged for review. 6. Click **🔄 Back-translate draft to English** — verify the meaning survived without needing to read Swahili. 7. Click **📄 Export target-language PDF** (or **Export as Markdown**) — get a reviewer handout. 8. **Now switch to Akuapem Twi** and run the *same* passage. Output quality drops noticeably — this is the **low-resource stress case**, and it shows exactly why the grounding + back-translation + native-speaker-review guardrails exist. --- ## 🔴 AMD Compute Usage This is **real GPU inference on AMD hardware**, configured end-to-end on the AMD Developer Cloud — **not a hosted-API wrapper**. The full trail (notebook, logs, `rocm-smi` output, benchmarks, generation samples) is committed and reproducible in **`evidence/amd-gpu/`**. | | | | --- | --- | | **Model** | Google **Gemma 3 12B Instruct** (`google/gemma …