Gemini Research Assistant

Inspiration

It started with a math exam I had less than one night to prepare for. I had a pile of topics to cram and no fast way to turn them into something I could actually study from — just raw material and a ticking clock. That frustration was the spark: what if you could feed in a document and immediately get back exactly what you need to study? I built ScholarAI so nobody else has to pull an all-nighter re-reading the same paragraph five times — and so it's useful far beyond math, for any subject where someone just needs to turn dense material into something they can actually learn from.

What it does

ScholarAI turns any academic document (PDF, DOCX, PPTX, XLSX, images, and more) into structured study material:

Instant analysis — a clear summary, key ideas, and a simplified explanation of the material
Auto-generated flashcards for active recall
Drake, an AI study assistant, answers follow-up questions grounded in your uploaded documents
Multi-language support (English, Russian, Korean) with on-the-fly translation of results
Usage-based plans (Free / Pro / Premium) so the app scales from a single student to heavier daily use

How we built it

Next.js 16 (App Router, Turbopack) for the frontend and API routes
Supabase for authentication and Postgres storage
Google Gemini API for document analysis and the Drake chat assistant
Stripe for subscription billing
Deployed on Vercel, with the whole pipeline — upload, extraction, analysis, chat — running through serverless functions

Challenges we ran into

Two stood out. First, Google deprecated the Gemini model we originally built against mid-way through — every analysis request started failing in production with no warning, so we had to migrate to a self-updating model alias to make sure the app doesn't silently break again the next time a model gets retired. Second, tracking down why a fix that worked locally still failed in production taught us a lot about how Vercel deployments, aliasing, and caching actually work under the hood — not something you usually learn until it bites you during a live demo.

Accomplishments that we're proud of

Getting the full pipeline — upload, AI analysis, chat, translation, billing — working end to end as a small team, and building real production-debugging skills (reading deploy logs, tracing errors back to root cause) along the way.

What we learned

That "it works on my machine" and "it works in production" are two different claims — and that AI APIs move fast enough that you have to design for model deprecation from day one, not as an afterthought.