Case Study Overview

People living with diabetes have health questions every day, and general AI chatbots answer from memory, sometimes confidently wrong. Sakkar AI was built as a daily diabetes companion that looks up published research first, drawing on 1.4 million PubMed articles, and answers only from what it finds, so every answer can be checked.

Challenges

Health is a field where a fluent but wrong answer does real harm. Answers had to come from real studies, not the model's memory. The assistant had to find the right research for questions asked in everyday language, show where each answer came from, and say clearly when the research did not cover a question. All of this had to work in one product across mobile and web.

Solution

We built Sakkar AI from scratch as a retrieval-based (RAG) assistant:

• Search over 1.4 million PubMed articles for every question
• Retrieval that passes only the most relevant studies to the AI model
• Answers written from the retrieved studies, with citations
• Daily commitments for healthy habits like walking, meditation, sleep and water
• A mobile app and a web app running on one shared backend

Final Deliverables

Implementation Details

• Mobile app built in React Native, one codebase for the phone experience
• Node.js and Express API on AWS that handles search, retrieval and answer generation
• Supabase for user accounts and app data
• A retrieval pipeline that turns a plain-language question into a research search, then sends only the best matches to the language model
• Prompts and rules that require a source behind every answer

Results

• Answers grounded in research: every reply is built from retrieved studies, not the model's memory.
• Sources users can check: citations come with each answer, so users can read the original study.
• One product, two platforms: mobile and web run on a single shared backend.
• Built end to end: our team delivered the app, backend, AI pipeline and cloud setup.

Sakkar AI shows what a focused RAG product looks like in a high-stakes field: a mobile and web assistant that answers from published research, shows its sources and does not guess.