Erica AI | Devpost

Inspiration

It’s 10:00 PM on a Sunday. You have a Computer Science exam at 8:00 AM. You have 400 pages of slides, three messy Notion pages, and a practice test you don't understand. The temptation? Plug it all into an LLM and ask for the answers. The problem: You haven't actually learned anything, and tomorrow morning, that's going to show.

We realized that AI is currently being used as a crutch rather than a coach. We wanted to build something that doesn't just give you the "what," but masters the "how." We were inspired to turn the "panic-study" session into a structured, scientifically-backed learning cycle.

What it does

Erica is a personal AI study agent. She ingests your raw chaos, PDFs, scribbled notes, and practice exams, and filters them through a tailored 5-stage learning framework:

Experience: Engaging with the core material. Reflect: Identifying personal gaps in knowledge. Internalize: Connecting new info to what you already know. Conceptualize: Breaking down complex theories into mental models. Apply: Real-world problem-solving.

For students, it’s a personalized roadmap that respects their time and brain type. For teachers, Erica provides a "heat map" of student understanding, showing exactly where the class is tripping up so they can adjust their lectures in real-time.

How we built it

Frontend: A slick and responsive UI built with Next.js for seamless state management. Backend: A Flask 3.x API for application logic and AI orchestration. Database: Supabase for our primary cloud storage and SQLite for rapid local prototyping and indexing. The "Brain": Gemini handles the deep content analysis and pedagogical structuring, while ElevenLabs provides the auditory layer for students who learn better by listening. Processing: We used PyPDF and custom Python scripts to scrape and clean messy academic documents into AI-ready data.

Challenges we ran into

The "PDF Nightmare" was our first big wall. Academic resources are messy, professors use different slide formats, handwritten notes are often unreadable, and practice tests are usually poorly formatted. Getting PyPDF to extract clean, usable data for Gemini without losing the context of diagrams or formulas was a constant battle.

Beyond the technical, the biggest hurdle was coming up with coaching. Standard AI just wants to give you the answer. We had to spend hours fine-tuning our prompts to ensure Erica acts as a Socratic coach, not a shortcut. If a student has ADHD or a specific learning disability, they don’t need the answer faster; they need the concepts broken down into smaller, high-stimulation milestones. Balancing deep personalization without compromising backend performance was a major challenge.

Accomplishments that we're proud of

We are most proud of the MANIM animation generation for mathematical based questions that need explaining. We are also proud of completing the full student flow which integrates our sparring partner.

What we learned

We learned that it’s not just about feeding the AI data, but about grounding it properly so it doesn’t hallucinate a fake formula during a midnight study session.

We also realized that UX for education is different. Students are stressed, tired, and often overwhelmed. We learned that the UI needs to feel calm, with less clutter and more focus. This project forced us to step out of our "coder brains" and think like a student who is three minutes away from a breakdown. We realized that a well-structured learning path is more valuable than a thousand "quick answers."

What's next for Erica AI

Canvas, Moodle, & Blackboard Integration: We want Erica to live where the students are. No more manual upload, just sync your course and your study plan is ready.

Collaborative "Study Rooms": Letting Erica facilitate group study sessions where she can identify which student understands a topic best and "nudge" them to explain it to their peers. Welcome to Erica, an Ai learning assistant for every student. Erica solves the lack of personalized learning present in America today through curated lesson plans, a back and forth Q&A dialogue, and finally a comprehensive test.