Hanoi, 28/06/2026

DEMO DAY

Virtual AI Teaching Assistant

Hồ Thành Tiến, Nguyễn Thái Bảo, Trần Mạnh Chánh Quân

​

Mentor: Đặng Hải Lộc

​

Agentic RAG · Zoom Bot · 24/7

Reduce TA workload by ≥70% with student response times under 10 seconds.

CONTEXT

Where the Problem Begins

How a single live MOOC session turns support into a bottleneck.

🎓

Massive Live Class

One MOOC session runs with thousands of learners online at the same time.

💬

Questions Flood In

Hundreds of questions pour into Zoom Chat and the course forum throughout the session.

📝

Manual TA Work

TAs must read, categorize, search for materials, and answer each question one by one.

⚡

Learners Can't Wait

Students need answers instantly to keep their learning momentum unbroken.

The breaking point — as enrollment grows into the thousands, a fully human-driven support model can no longer scale.

PROBLEMS

Students and Teaching Assistants Are Overwhelmed

⏳

Slow Response Time

Students wait hours for TA responses, disrupting their learning momentum

🔁

Repeated Questions

~60% of questions are repetitive, forcing TAs to handle thousands of inquiries per course

📝

Manual Note-Taking

Students must take notes manually, distracting them from the lecture

Confirms the problem — he agrees that online classes often lack enough teaching assistants, and that instructors can get overloaded or absorbed in teaching, so chatbox questions get missed — demotivating learners and widening the gap with students.

​

Hà Minh Tuấn - Lecturer at the Faculty of Mathematics and Computer Science, VNUHCM - University of Science (HCMUS)

GOALS

Virtual AI Teaching Assistant

MOOC students and teaching assistants are overwhelmed by a high volume of questions and slow response times. We developed an AI Teaching Assistant platform for large-scale online courses (MOOCs) that automates learning support using AI, reduces TA workload, shortens response time to under 10 seconds, and enhances the learning experience through question answering, AI-generated notes, mind map generation, and learning progress analytics.

⚡ Response Time < 15 Seconds

📉 Reduce TA Workload by ≥70%

🕐 Smart Note-Taking & Review

Stack: FastAPI · Next.js 16 · PostgreSQL · Qdrant · OpenAI GPT-4o · Docker · GitHub Actions CI/CD

Direct Integration: Zoom Meeting Bot · YouTube Live

SOLUTION

Use case diagram

CORE FEATURE

3 Must-Have Features

01

🤖

AI Course Assistant

Agentic RAG

• Answer Questions from Course Materials with Source Citations
• ReAct Agent Loop: FAQ → Vector Search → SQL → Web → TA
• Automatically Detect Out-of-Scope Questions and Escalate to TAs
• Detect Duplicate Questions and Respond Using the FAQ Cache
• Maintain Conversation History for Each Zoom Thread

02

📋

AI Lecture Notes

Automatic Summarization

• Automatically Summarize Each Lecture After Class
• Extract Key Concepts & Important Terms
• Generate Well-Structured, Review-Ready Notes

03

🗺️

Mindmap Generator

Knowledge Visualization

• Generate Mind Maps from Course Content
• Support Knowledge Review and Retention
• Export in Mermaid / Markmap Format

SYSTEM ARCHITECTURE

K3s · CI/CD · Cloudflare · Observability

TECHNICAL ARCHITECTURE

Agentic RAG Pipeline

AGENT TOOL CALLING & KNOWLEDGE

TECH STACK

RAG & PIPELINE CONFIDENCE

01

📥

Indexing

Offline

• Ingest PDF, DOCX, PPTX, TXT files
• Semantic chunking
• Build embeddings & BM25 index
• Store vectors in Qdrant

02

🔎

Retrieval

Query-time

• Embed user query
• Hybrid Retrieval (Dense + BM25)
• Rank with Reciprocal Rank Fusion
• Dedupe & filter by course
• Select Top-K context

03

✨

Generation

GPT-4o

• Generate answers with GPT-4o-mini
• Answer only from course materials
• Cite sources on every answer
• Reject unsupported queries

04

📚

Citation

Transparency

• Answers cited with filename & snippet
• Document traceability (doc_id)
• Context-grounded generation
• Fallback prevents hallucinations

05

🔁

Duplicate Detection

Cache

• Embedding-based duplicate detection
• FAQ cache (Similarity ≥ 0.80)
• Redis cache with TTL
• Jaccard deduplication

06

📊

Impact

Efficiency

• 100% source-cited responses
• Up to 60% fewer LLM calls
• Faster repeat answers
• Fewer hallucinations
• Lower cost via caching

MONITORING & OBSERVABILITY

01

📊

Metrics

Prometheus

• Latency, throughput, error rate
• OpenAI cost tracking
• Cache hit / miss ratio

02

📝

Logging

Grafana Loki

• Structured JSON logs
• DEBUG / INFO / WARN / ERROR
• Centralized log aggregation

03

🚨

Tracing

Jaeger

• Distributed tracing
• Trace ID / Span ID
• End-to-end request flow
• Performance bottleneck analysis

 RAG Quality (RAGAS)

Performance

• Faithfulness: 0.89
• Answer Relevancy: 0.83
• Context Precision: 0.79
• Context Recall: 0.8

• P95 Latency: 15s
• Error Rate: 1%
• Answer Rate: > 95%
• Token cost/query tracked

PRODUCT DEMO

Results & Conclusion

Successfully deployed and operated during a Zoom class session.

< 15s

Average Response Time

Compared to the previous hours

≥ 70%

Reduced Teaching Assistant Workload

Core Design Objective

24/7

24/7 Continuous Operation

No Human Supervision Required

0.70+

RAGAS Faithfulness Score

Automated CI/CD Pipeline

The AI Virtual Teaching Assistant addresses the issue of tutor overload and significantly improves the speed of learning support for students in large-scale online classes.

“ The solution is a seamless combination of Artificial Intelligence and Human expertise: AI automatically handles common questions in real time, while teaching assistants focus their expertise on more complex inquiries or those requiring high reliability verification ”

**Thank you for your attention!**

🤖

Agentic RAG

ReAct Loop

⚡

< 15 Seconds Response Time

📉

≥ 70% Teaching Assistant Workload Reduction

**MOOC learners and teaching assistants deserve better support—and AI can deliver it today.**

Team 03 · AI20K-019 · VinUni / Vingroup · GitHub: C2-App-003 · Demo: https://staging-c2-app-003.khoav4.com/

Q & A