AfterProof

AfterProof verifies learning, not just submissions using AI-powered oral defenses to ensure students truly understand their work

Comment

Inspiration

In the age of generative AI, written assignments are no longer a reliable signal of learning. Students can produce high-quality submissions in seconds—but that doesn’t mean they understand them.

We realized the real problem isn’t AI usage—it’s verification of understanding.

So we flipped the question: Instead of asking “Was this written by AI?”, we ask
“Can you defend it?”

That idea became AfterProof.

What it does

AfterProof transforms traditional assignments into interactive oral defenses powered by AI.

  • Students submit their work (PDF/Doc)
  • The system analyzes the submission using AI
  • It generates targeted, submission-specific questions
  • Students go through a live, voice-based knowledge check
  • The system evaluates their answers and generates a detailed report

This shifts evaluation from static grading to active understanding verification.

How we built it

We designed AfterProof as a multi-service AI system:

  • Frontend: React + Vite for instructor dashboards, student flows, and live sessions
  • Backend: Node.js + Express + Socket.IO for real-time orchestration
  • Agent Service: Python + Flask powered by Vertex AI Gemini

Core AI Pipeline:

  1. Analysis Agent → extracts concepts, claims, weak points
  2. Question Designer Agent → generates targeted questions with follow-ups
  3. Evaluator Agent → scores understanding based on responses

We integrated:

  • Google Cloud Speech-to-Text & Text-to-Speech for live interaction
  • Firestore for real-time state management
  • Cloud Storage for submissions
  • Cloud Run for scalable deployment

Challenges we ran into

  • Making AI outputs truly grounded
    Preventing generic questions and ensuring they reference the actual submission

  • Real-time system complexity
    Coordinating audio, transcription, sockets, and AI responses seamlessly

  • Evaluating understanding vs correctness
    Designing prompts that detect depth of knowledge, not just “right answers”

  • Handling vague responses
    Building adaptive follow-ups that push students to clarify and justify

Accomplishments that we're proud of

  • Built a fully working end-to-end AI system across multiple services
  • Implemented a multi-agent reasoning pipeline (analysis → questioning → evaluation)
  • Enabled real-time oral defenses with live AI interaction
  • Successfully deployed a production-ready system on Google Cloud
  • Created a solution that redefines how learning can be assessed in the AI era

What we learned

  • Detecting AI usage is a losing battle—verifying understanding is the real solution
  • Prompt engineering is critical for controlling AI behavior in complex workflows
  • Real-time AI systems require strong backend orchestration
  • Even simple questions can expose deep gaps in understanding
  • The best AI products combine UX, system design, and intelligent reasoning

What's next for AfterProof

  • Full authentication and classroom management system
  • Session recording and playback for review
  • More transparent and explainable evaluation metrics
  • Support for more domains and assignment types
  • Enhanced AI tutor for personalized feedback and learning

Final Thought

In a world where AI can generate answers,
the true signal of learning is the ability to explain them.

AfterProof makes that measurable.

Built With

Share this project:

Updates