TAura

Inspiration

Teaching assistants play a vital role in supporting students, but human TAs can’t always be available 24/7 - and scaling personal help across hundreds of students is hard. We wanted to build an AI teaching assistant that is course-aware, policy-conscious, and human-like in tone, giving every student fast, trustworthy guidance while empowering instructors to stay in control.

What it does

Ingests each course’s syllabus, slides, assignments, and transcripts to create a scoped knowledge base.
Provides students with a chat interface to ask course questions and receive cited, policy-compliant answers.
Gives instructors a dashboard to upload materials, define policies, select a TA persona, view analytics, and test responses.
Automatically extracts structured syllabus info (office hours, late policy, etc.) for quick student answers.
Generates lecture summaries, study packs, and coverage checks to ensure all materials are aligned with student needs.

How we built it

Frontend: React + Tailwind for the instructor dashboard and student chat UI.
Backend: FastAPI for APIs, authentication, and course/material management.
Ingestion pipeline: Converts PDFs → text → chunks → embeddings (via LangChain + OpenAI embeddings) and stores them per course_id.
Retrieval: RAG system filters relevant chunks per course context.
LLM service: Claude 3.5 for grounded answers with in-context citations and prompt-level policy enforcement.
Database: PostgreSQL for users, courses, policies, and logs.
Storage: AWS S3 or Supabase for file management.

Challenges we ran into

Building a scoped retrieval system that avoids content leakage across courses.
Designing a prompt policy framework that prevents full-solution or exam-related answers while still being helpful.
Achieving a natural, human-like “TA” tone that adapts between strict and friendly modes.
Managing long documents and embedding chunk optimization for lecture transcripts.
Balancing answer latency and accuracy during RAG retrieval.

Accomplishments that we're proud of

Built a fully functional multi-role platform (instructor + student) within hackathon constraints.
Implemented policy-aware prompting and auto-cited RAG answers.
Developed a TA Simulator Console to preview how the AI would answer student questions before publishing.
Created a Coverage Check that identifies missing or outdated materials based on student questions.
Generated lecture summarization and study packs automatically for uploaded lectures.

What we learned

The importance of structured content ingestion for reliable course-specific retrieval.
How prompt design and role conditioning drastically affect trust and tone in educational AI.
That even simple analytics (like “most asked questions”) give instructors deep insights into student struggles.
Building an AI that’s aligned with academic integrity is as much about policy and UX as model performance.

What's next for TAura

Add Canvas / Blackboard integration for direct LMS connection.
Deploy a TAura Analytics Dashboard that visualizes student engagement and friction points.
Implement multimodal support for lecture slides (vision-text embedding).
Expand persona packs (e.g., Policy-First TA, Friendly Explainer, Tutor Mode).
Release a public instructor onboarding flow for any educator to create their own course assistant.