MAM-AI

MAM-AI is a smart search application developed for nurses and midwives in Zanzibar. This repository is an MVP for this, as submitted to the Gemma3n Kaggle challenge.

There are three main folders: rag, app, and finetune. rag is used for document preprocessing. app contains the actual app. finetune is the Gemma3n finetune which we sadly did not manage to deploy into the Android app (yet).

Install instructions

Under the GitHub releases tab there is an apk to install :) It may not work on an emulator (according to the Google AI Edge RAG documentation) since the underlying inference library needs real hardware.

You can also build it yourself with flutter build apk in the app/ directory.

Reproduction instructions

This is a rough sketch of how you could reproduce what we created in this project.

Prepare RAG docs:

1. Curate documents you want to include (or use the Google Drive link from the writeup)
2. Run MMORE over these documents to extract their text
3. Chunk the documents using the scripts in rag/
4. Copy the chunks to the mamai_trim.txt in the assets folder of the Android app and uncomment the memorizeChunks() call
5. Run the app and wait for it to incorporate the chunks into the sqlite db
6. Re-comment the memorizeChunks
7. Use adb to pull the embeddings.sqlite for redistribution

How we developed the app:

• Flutter frontend with regular Flutter <-> Android FFI/bridging
  • Built this out to meet our needs
• Adapted the Google AI Edge RAG for the Android backend which runs the LLM

Serving the remote files to the users:

• Start an nginx server with a self-signed cert and the files (see below) in /var/www/html/

Finetuning (not included in app):

• Finetuning Dataset
• Finetuned Model
• Set Up a Python 3.10 Virtual Environment:
  • python3.10 -m venv .venv
  • Windows: .venv\Scripts\activate Linux: source .venv/bin/activate
• Install core dependencies
  • pip install -r requirements.txt
• Run the Training Script
  • python train.py

Remote resources

Note: because of the requirements to accept the Gemma3n license before using it, we do not provide the model files in this repo. Hence, we host everything on a temporary VPS which will only remain up during judging. If you are uncomfortable with this, you can simply replace the link with your own endpoint.

We host copies of various models on a temporary VPS. The app fetches these the first time it launches. This is just done for simplicity's sake.

They are as follows:

• Gecko_1024_quant.tflite: embedding model from litert-community/Gecko-110m-en
• sentencepiece.model: tokenizer from litert-community/Gecko-110m-en
• gemma-3n-E4B-it-int4.task: Gemma3n E4B
• embeddings.sqlite: pre-computed document embeddings