StudyPilot AI

Inspiration

As an engineering student, I spend hours turning dense lecture slides and PDFs into useful study materials. I rewrite notes, create flashcards, and speculate about potential exam questions. This process is repetitive and the same for every student, no matter the subject. I wanted to see if I could automate this entire workflow: upload once and instantly receive a complete study system.

What it does

StudyPilot AI takes any lecture document, whether slides, notes, or a PDF, and automatically generates a structured summary, flashcards, and a graded quiz from a single upload. Additionally, it functions as an adaptive tutor. A student can ask it to "teach me this like I'm 12" or "give me harder questions," and it will adjust its explanations or create new questions on the spot. As students take quizzes, StudyPilot tracks their performance by topic. It highlights weak areas, helping students know what to review before an exam instead of rereading everything.

How we built it

The project runs on Next.js 14 with the App Router, using Firebase for authentication and Firestore for storing documents, flashcards, quizzes, and progress history. At the core of the product is a single, well-structured prompt to Gemini 3.5 Flash that returns a summary, flashcards, and a quiz as one structured JSON object in a single call. This design keeps latency low instead of requiring multiple requests. The adaptive tutoring feature allows us to pass a student's natural-language request alongside the original document content directly to the model, rather than creating separate rules for different student inquiries.

We collaborated with Codex, powered by GPT-5.6, as an active build partner during the project. This included setting up the Next.js and Firebase structure, refining the JSON prompt schema until the output was consistently parseable, and building the Firestore aggregation logic for the weak-topics dashboard.

Challenges we ran into

Getting the model to reliably return clean, parseable structured JSON for the summary, flashcards, and quiz in a single call took significant iterations. Early versions produced inconsistent formats that broke the user interface, so we focused on refining the prompt and adding validation. We also made a conscious decision early on to cut live audio transcription from the MVP. This choice helped us keep the core pipeline (document to structured study material) solid for the demo instead of spreading our efforts too thin across multiple input types.

What we learned

Letting the model respond to open-ended instructions directly, such as "teach me like I'm 12," proved to be simpler and more effective than trying to categorize what a student might ask. We also discovered that the outcome of a hackathon relies heavily on focusing tightly on one effective pipeline rather than creating many superficial features.

What's next for StudyPilot AI

The next steps include supporting live audio and lecture recording transcription, creating a teacher dashboard to aggregate weak topics across an entire class, and predicting exam questions based on the topics that appear most often in a student's uploaded materials.

Built With

• adaptive-learning
• codex
• edtech
• education-technology
• firebase
• firebase-authentication
• firestore
• gemini
• gemini-3.5-flash
• generative-ai
• google-ai
• gpt-5.6
• javascript
• json
• llm
• nextjs
• node.js
• pdf-parsing
• prompt-engineering
• react
• rest-api
• tailwindcss
• typescript
• vercel