GSoC 2026 Interest: Gemini-API–Powered Intelligent Care Assistant
Yuyao XU
Hi Supreeth and Atharva,
I'm Yuyao Xu, a first-year MS CS student at Northeastern University. I've been looking through Catrobat's GSoC 2026 ideas and the Gemini-API–Powered Intelligent Care Assistant really caught my attention.
The responsible AI angle is what draws me most — building something that augments caregivers rather than replacing their judgment. I ran into similar constraints in a recent project: a Medical RAG Chatbot trained on 11,000+ real medical Q&A pairs, using hybrid BM25 + SBERT retrieval, FAISS indexing, and Flan-T5 / LLaMA-3.2 generation, evaluated with BLEU, ROUGE-L, and BERTScore. The key design decision was to ground responses strictly in retrieved evidence, not let the model generate freely — reliable and scoped, not a creative AI. That's the same discipline this project calls for, and I find it more interesting to work within than an open-ended research setting. Live on Hugging Face Spaces: https://huggingface.co/spaces/yyx11/diabetes-chatbot.
Before I put together a proposal, I had a few questions:
1. Is there a defined entry task for this project, or is that left open to the applicant?
2. Is there mock IoT data in the codebase to work with, or would I need to simulate that myself?
3. Any preference on Gemini API model tier or SDK version?
4. Would it be helpful to share a draft proposal with you for early feedback before the official submission?
Thanks for your time — looking forward to hearing from you.
Yuyao Xu
MS CS, Northeastern University
GitHub: https://github.com/yuyaoxu11-bit/Medical-RAG-Chatbot
LinkedIn: https://www.linkedin.com/in/yuyao-xu-498168270
Yuyao XU
Hello mentors,
I’d like to briefly share the pipeline direction I've been exploring.
The architecture I have in mind consists of five layers: IoT signal ingestion, routine summarization (using time windows), deviation detection, Gemini-based natural language explanations, and a human-in-the-loop step. My main goal is to keep the AI in an interpreter role only. I want to make sure there's always a supervisor check before any alerts actually reach the end user, avoiding any direct diagnostic inference.
For the prototype, I’ve found some solid public datasets that cover the signals mentioned in the project description (including timestamped location/activity events, sleep state sequences, physiological readings, and clinical event labels). I think these should work well to simulate real-world IoT streams for now.
That said, I'd like to ask the mentors:
Is there a specific reference dataset or schema the team prefers me to align with?
If simulated data is fine for the prototype, are there certain signal types or formats you’d like me to prioritize?
I'm happy to dive deeper into any of these points if you're interested. Thanks for your time!
Best,
Yuyao