Virtual Teaching Assistant (VTA)

VTA is your 24/7 AI study buddy where you can ask anything, solve problems, quiz yourself, learn with flashcards, and follow smart study plans. Personalized learning, always ready to help!

Comment

Inspiration

Office hours are limited, and not all students get equal access. Some face schedule conflicts with lectures, while others wait too long only to run out of time. Often, one student’s complex question consumes most of the TA’s bandwidth, leaving others without help. We wanted to build a Virtual Teaching Assistant (VTA) that’s always available, offers equitable support, and delivers AI-powered academic assistance anytime, no queues, no time restrictions.

What it does

VTA is an AI-powered platform that acts like a real TA but smarter, faster, and always online. It allows students to: • Ask Questions: Get instant, detailed explanations • Practice Problems: Solve problems with step-by-step feedback • Take Quizzes: Test understanding with AI-generated assessments • Flashcards: Learn via spaced repetition • Study Plans: Personalized AI study schedules

How we built it

Frontend: Built using Luvable, a no-code UI platform for rapid interface development Backend: Developed using LangGraph and LangChain, powering node-based flows for chat, quiz generation, and task routing Model: We fine-tuned the meta-llama/Llama-3.2-3B-Instruct model using the unsloth library with LoRA adapters for efficient deployment

Challenges we ran into

• Model–Backend integration: With different teammates working on the backend and model finetuning, merging both into LangGraph’s node structure was difficult • LoRA compatibility: Ensuring that the fine-tuned model worked in 4-bit precision inside our orchestrated flows took debugging • Syncing team deliverables: Aligning the timing and format across frontend, backend, and model workstreams under time pressure was a real challenge

Accomplishments that we're proud of

• Built a fully working AI Teaching Assistant with five core features
• Successfully fine-tuned and deployed a LoRA model using unsloth
• Created an intuitive and beautiful UI with Luvable, all without writing CSS
• Made a modular backend pipeline using LangGraph for future extensibility

What we learned

• How to use LangGraph and LangChain to build agentic workflows
• Efficient LLM finetuning using LoRA + Unsloth in 4-bit memory
• How to rapidly develop frontend UIs using no-code tools like Luvable
• Real-time collaboration across model, backend, and UI layers

What's next for Virtual Teaching Assistant (VTA)

Integrating the frontend with the backend Voice-based Q&A using Whisper and Gemini Multimodal support (upload a photo of a problem to get an answer) Finetune on different subject materials

Built With

  • huggingface
  • langchain-?-frontend:-luvable-(no-code-builder)-?-model-tools:-hugging-face-transformers
  • langgraph
  • llm-&-finetuning:-meta-llama-3.2?3b
  • lora-?-backend:-python
  • lovable
  • peft
  • peft-?-data:-in-memory-json-storage-?-(planned):-whisper-(voice)
  • python
  • typescript
  • unsloth
Share this project:

Updates