Home Page of PAAL
Tap-based tutoring for USF courses: choose chapter & topic, get grounded explanations, summaries, prompts, and notes PDF
Generate course-specific quizzes by difficulty/type, submit answers, get graded feedback, and track your score history
See progress insights from your recent chats and quiz scores: strengths, focus areas, recommendations, and trends.

PAAL — Promptless AI Assisted Learning

Inspiration

USF students don't need another generic chatbot—they need help that matches their syllabus, their textbook, and their deadline. Most students aren't prompt engineers: fighting a blank ChatGPT box to "ask the right question" is friction, not studying.

We built PAAL (Promptless AI Assisted Learning) so learners can get instant, course-specific support—grounded in real course materials—without turning study time into prompt engineering time.

What It Does

PAAL is built around a promptless, tap-first workflow:

Teach / Explain — Start from course + chapter/topic and launch a structured explanation—no essay prompt required.
Practice Quiz — Generate assessments aligned to what you're actually studying, with configurable formats and difficulty.
Summarize & Study Outputs — Turn the session into summaries and printable study notes so the value isn't trapped in an endless chat thread.

Behind the calm UI, every action maps to a deterministic "learning job" on the backend—so students click to learn, not type to beg.

How We Built It

Layer	Technology
Frontend	React 19 + Vite + React Router + Tailwind CSS
Auth	Clerk (`@clerk/clerk-react`) + JWT-backed API calls
Backend	FastAPI + SQLAlchemy + SQLite
AI System	CrewAI + Google Gemini (Tutor, Assessment & Analytics agents)
Grounding / RAG	Pinecone vector DB + course PDF indexing

Frontend: Multi-route app (Study Hub, Quiz, Analytics, auth screens) with Markdown rendering and PDF export utilities.
Auth: Clerk JWT verification on the backend keeping the frontend token path reliable across routes and reloads.
Backend: FastAPI exposing /api/chat plus course catalog/outline endpoints; CORS configured for the Vite dev origin.
AI System: CrewAI orchestrates a sequential multi-agent crew powered by Google Gemini.
Grounding / RAG: Retrieved textbook passages injected with strict course filters and explicit guardrails—reducing hallucinations compared to open-internet chatbots.

Challenges We Ran Into

End-to-end integration — Wiring Vite ↔ FastAPI with CORS, proxies, and credentialed API calls while keeping local dev smooth.
Auth + API security — Placing Clerk JWT verification on the backend and keeping the frontend token path reliable across routes and reloads.
Agent output quality — CrewAI's verbose logging can introduce formatting noise—post-processing was needed so downstream UI/Markdown consumers get clean, parseable responses.
Grounding at scale — Making retrieval deterministic (exact course metadata filtering + injected context blocks) so RAG doesn't depend on an LLM "remembering" to call tools correctly.
Product truth vs AI hype — Keeping the UX promptless while still supporting deeper follow-ups—balancing simplicity with real tutoring depth.

Accomplishments That We're Proud Of

We shipped a full-stack, multi-agent AI application with:

[x] Real authentication with Clerk + JWT
[x] Vector-backed RAG grounding with Pinecone
[x] Multi-agent AI crew (Tutor, Assessment, Analytics)
[x] Polished student UX with multi-route React app
[x] End-to-end working system in a compressed hackathon timeline

PAAL isn't a slide-deck demo: it's a working system with routes, persistence, and an agentic backend wired together.

What We Learned

We leveled up on:

Full-stack integration — React SPA + FastAPI + Auth + env management
Agentic AI design — Turning "chat" into task contracts
RAG as a reliability layer — Not just for knowledge, but for reducing hallucinations
Student-first UX — Building interfaces that respect how students actually study