Mindsieve interface

🧠 MindSieve: Gemini‑Powered Tutor with Elastic Recall

Inspiration

Learning about Computer Science and finding relevant research papers has been fairly difficult in the past.

MindSieve started as a way to make technical ideas feel approachable again --- a place where students and lifelong learners can ask "how does this work?" and get a clear, sourced explanation.

It uses AI not to replace curiosity, but to guide it --- connecting what you already know to what experts have written.

What it does

MindSieve acts as a computer science tutor that explains concepts, summarizes key ideas, and links you to deeper reading.

It utilises the power of Google Gemini AI and Elasticsearch to find up-to-date research papers.

Ask it about data structures, neural networks, or compression --- it answers with short, structured sections that build from beginner to expert insight.

Each response ends with suggested follow‑up topics and links to real papers, giving students a safe launch point for exploring further.

It's not a chatbot that lectures --- it's a study companion that helps you find your own "aha!" moments.

How we built it

Frontend: Built with Next.js, Tailwind, and shadcn/ui for clarity and smooth user flow.
Search: ElasticSearch combines keyword (BM25) and semantic (vector) search for concept‑level matching.
AI layer: Vertex AI's Gemini provides concise, Markdown‑formatted explanations and follow‑up ideas.
Embeddings: text‑embedding‑005 vectors make the retrieval contextual, based on a large set of computer science abstracts.
Automation: Cloud Run and Cloud Scheduler handle regular data refreshes and background indexing.

Challenges we ran into

Finding the balance between simplicity and accuracy --- keeping explanations short without losing depth.
Making Gemini consistently follow a tutor‑style structure instead of a long essay.
Keeping search relevant while avoiding too much technical jargon from raw papers.
Creating a smooth streaming interface that feels natural to interact with.

Accomplishments that we're proud of

Designed a tutor experience that feels friendly, not overwhelming.
Built a working hybrid retrieval pipeline combining Elastic and Vertex AI in real time.\
Kept everything serverless and fast enough for everyday learners.
Focused on clarity, learning, and transparency rather than flashy AI tricks.

What we learned

The quality of an answer often depends more on structure and tone than raw model power.
Retrieval grounding keeps AI honest --- students appreciate seeing where ideas come from.
Good educational AI should encourage exploration, not end it.
Simplicity in design goes a long way toward keeping people engaged.
Setting up the ElasticSearch Agent on Google Cloud using a custom VM and experimenting with the technology