First public build of Remedium for The Gemma 4 Good Hackathon submission.
Setup
You need two files: the APK (attached below) and the Gemma 4 E2B model
(download separately from Kaggle).
1. Download the Gemma 4 E2B model
Get gemma-4-E2B-it.litertlm (2.59 GB) from Kaggle:
https://www.kaggle.com/models/google/gemma-4
The model file is too large for GitHub releases (2 GB hard limit), so it
must be downloaded directly from Google's official source.
2. Push the model to your phone
With your phone connected via USB and ADB enabled:
adb push gemma-4-E2B-it.litertlm /data/local/tmp/gemma.litertlm
Or use the included scripts:
- Linux/macOS:
chmod +x setup.sh && ./setup.sh - Windows:
setup.bat
3. Install the APK
adb install remedium-v1.0.apk
Or transfer the APK to your phone and tap to install (enable "Install from
unknown sources" in Settings first).
Grant camera and microphone permissions on first launch.
What works
- On-device OCR (ML Kit) on 28 verified Indian generics + 7,477 brand fallback
- Tier-based trust UI (verified / identified / unknown - chatbot only on verified)
- Gemma 4 E2B grounded Q&A in English and Hindi
- Drug interaction reasoning across two scanned medicines
- Safety refusals with arithmetic on overdose questions
- Bilingual TTS readout
- Fully offline - no network calls after setup
Tested on
POCO X6 5G - Snapdragon 7s Gen 2, 11.5 GB RAM, Android 16.
Minimum recommended: 8 GB RAM phone, Android 10+, arm64-v8a.
Safety
24/25 on our adversarial grounding test (10/10 on safety-critical refusals
including overdose, self-harm, and diagnosis requests). The full eval CSV
is in the repo at app/src/test/resources/grounding_eval.csv.
See README for an honest list of known limitations.