JemmaPass is a real, working Android app — not a mockup. This page is your single entry point: watch the 3-minute demo, inspect the 352+ Kotlin source files, download the verified APK and run it on a Pixel 9, then audit every clinical claim with primary-source citations.
What it is, why it matters, who it serves. Read this first.
YouTube embed + scene-by-scene storyboard with the Shikoku earthquake scenario.
Kurodo, Haru, Kamekichi — their real IPS profiles in _j 1.2 format.
352+ Kotlin files. Apache 2.0 dependencies. Reproducible build with NDK r27.
Direct APK + SHA-256 + install instructions for your Pixel 9 (or any Android 15 device).
The official submission on Kaggle, with all evaluation rubric criteria addressed.
The Gemma 4 Good Hackathon judges on three criteria. Here is how we map our submission to each, with direct links to verify.
Does your project address a significant real-world problem? Is the vision inspiring?
Is the video exciting, engaging, well-produced? Does it tell a powerful story?
Is the tech innovative, functional, well-engineered? Not just faked for the demo?
widgetCategory="home_screen|keyguard") — double-tap SOS, zero unlock required.Per Section 2.5 of the Competition Rules, projects are eligible for both a Main Track Prize AND a Special Technology Prize. JemmaPass is positioned for maximum overlap with the rubric.
Air-gap clinical decision support is a category-creating play. No existing project ships a Gemma 4 + IPS R4 + Nearby Mesh fusion. Real lives-saved potential in disaster zones.
First fusion of HL7 IPS R4 + DDInter 2.0 + Gemma 4 + Nearby. Zero-translation paradigm: clinical codes resolved deterministically via embedded multilingual KB, not LLM translation. Hospital-grade accuracy.
Born from Shikoku pilgrimage × 3 and the Noto 2024 earthquake volunteer experience. Offline edge-based response is the entire premise. Works on a mountain when cell towers collapse. Lock-screen widget triggers SOS in zero unlock — the victim never has to fumble with PIN or biometric.
15 UI languages, 25+ clinical-term languages, no subscription, no account, no internet ever required. Built for refugees, pilgrims, and grandmothers — not Silicon Valley.
Gemma 4 E2B (2.4 GB) and E4B (3.4 GB) both supported via LiteRT-LM 0.11.0 native function calling. 49 @Tool functions across 6 agentic surfaces. The deepest LiteRT integration in the field.
Maximum cumulative payout possible: $60,000 (Main 1st + one Special Technology Prize).
Per the hackathon Live Demo requirement, you can download the signed APK directly. The build is reproducible from the GitHub source per the README. We recommend a Pixel 9 (Android 15) for the 16 KB page-size verified experience.
jemmapass.net/models/. Allow ~10 minutes on a fast Wi-Fi.The rubric explicitly asks: "Is the technology real, functional, well-engineered, and not just faked for the demo?" Here is how to verify, end-to-end, that nothing in our submission is staged.
It's a real query against knowledge_full.db. Inspect JemmaTools.kt line ~340 (checkOneDrugAgainstFocusProfile). It executes a real SQL JOIN across ddi_atc_pairs and terminology_codes in under 50 ms.
No LLM translation — that would be slow and hallucination-prone. We do code-to-display lookup via ips_valuesets_translations.lang='ja-JP'. SQL query, deterministic, sub-millisecond. Inspect JemmaProfileHydrator.kt.
Yes. Strategy.P2P_CLUSTER uses BLE + Wi-Fi Direct radios — no internet involved. Put both phones in airplane mode and watch JemmaNearbySosService.kt still discover and relay. The 6 wire chunk types (V/VR/S/SR/E/F) handle direct + multi-hop relay.
Verify with adb shell ls -la /sdcard/Android/data/be.heyman.android.jemmapassdemo/files/knowledge-full-db/. File size: 3,360,727,040 bytes. SQLite PRAGMA integrity_check; returns OK.
We've cross-verified every claim against authoritative sources. No marketing inflation.
jemma_kb_audit.py against the production knowledge_full.db (3.36 GB · 58 tables) · see the full audit ↗
In September 2024, at the HL7 FHIR Connectathon, Dave deBronkart ("e-Patient Dave") proved patient-published IPS via QR was possible. His proof-of-concept (publicly viewable) was an inspiring milestone — but it requires the receiver to be online. JemmaPass inverts the architecture.
_j2:)Same standard (HL7 FHIR IPS R4). Different architecture. The QR code IS the patient summary — not a pointer to one.
Three companion pages with code excerpts, primary-source citations, and prize track alignment. Open the ones that match your evaluation angle.
The 5 layers of the Silicon Shell · real Kotlin code excerpts · 6 chunk types · deflate-raw codec.
302K DDI records explained · primary sources cited · the SHL comparison expanded.
5 tracks · why we fit each one · payout matrix · realistic combinations.
Four companion pages with build instructions, benchmark numbers, anticipated questions, and the human story behind the project.
Full toolchain pins · 6-step build · 16 KB ELF proof · APK signature verification · network isolation tests.
Real Pixel 9 metrics · <200ms red alert · <50ms DDI query · model load · mesh discovery · device matrix.
15+ honest answers to skeptical judge questions · technical · clinical · regulatory · limitations.
58 tables · 3 views · 88 indexes · 37.9M rows · 4 FTS5 · auto-generated from knowledge_full.db.
Origin story · 3 Shikoku pilgrimages (2016–2023) · Noto 2024 volunteer · acknowledgments.
MIT during the hackathon → relicensed to CC-BY 4.0 upon any prize award, per the Competition Rules.
Inherits CC BY-NC-SA 4.0 from DDInter 2.0. SNOMED CT IPS Free Set under CC BY 4.0. UMLS under Metathesaurus License.
Gemma 4 under Apache 2.0 (Google DeepMind, April 2026). LiteRT-LM 0.11.0 same terms.
Zero telemetry. Zero analytics. Zero cloud calls during clinical decisions. PHI never leaves the device by design.
Clinical disclaimer: JemmaPass is a decision-support tool for emergency resilience and patient empowerment. It is not a medical device under EU MDR / FDA SaMD classification, and does not substitute for professional medical judgment.
Three suggested paths depending on what you want to verify next.