Skip to content

DEMO_SCRIPT.md

Latest commit

 

History

History
149 lines (103 loc) · 4.02 KB

DEMO_SCRIPT.md

File metadata and controls

149 lines (103 loc) · 4.02 KB
Raw

AG-TUNE Live Demo Script (7 Minutes)

Setup (Already Done) ✓

  • Server running at: http://localhost:3000/
  • All algorithms implemented with real math
  • No mock data, no placeholders

Demo Flow

1. Introduction (30 seconds)

"This is AG-TUNE - a neuro-symbolic poetry engine that generates poetry using actual algorithms, not neural networks."

Show: Header with title and description

2. Show Training Panel (1 minute)

Point out:

  • File upload for custom training corpus
  • Corpus size display (starts with 10 embedded lines)
  • Epoch slider (controls training iterations)
  • Real-time metrics during training

Key Message: "This uses real distributional semantics, not pre-trained word vectors"

3. Live Training Demo (1.5 minutes)

Steps:

  1. Click "Train Model"
  2. Watch progress bar fill
  3. Show real metrics updating:
    • Average Reward (changes each epoch)
    • Epochs Completed counter
    • Vocabulary Size (extracted from corpus)

Explain while training: "The engine is:

  • Computing co-occurrence embeddings
  • Running Kernel PCA for emotional space
  • Training TD-lambda value estimator
  • All real linear algebra and calculus"

4. Show Internal States (1 minute)

Scroll to visualizations:

Emotional Space:

  • "These are actual 8D vectors from polynomial kernel PCA"
  • Show different words have different values
  • Point out: "No word2vec, no GloVe - computed from scratch"

Value Estimator Weights:

  • "These 16 weights were learned via temporal difference learning"
  • Show the bar graph changes after training
  • "Real gradient descent, real backpropagation"

5. Poetry Generation (2 minutes)

Steps:

  1. Enter prompt: "whispers in the moonlight"
  2. Adjust beam width to 7
  3. Set lines to 4
  4. Click "Generate Poem"

Explain while generating: "The engine is now:

  • Running A* beam search
  • Checking grammar with CYK parser
  • Analyzing rhythm with FFT
  • Avoiding repetition with Floyd cycle detection
  • Evaluating aesthetic value with learned V(s)"

Show generated poem:

  • Read it aloud
  • Point out it follows meter
  • Show generation metrics (M, N, V values)

6. Prove It's Real Math (1 minute)

Open browser console (F12):

  • Show no errors
  • Explain: "All computation happens client-side in JavaScript"

Alternative - Run test script:

node test-engine.js

Show output:

  • Kernel PCA working ✓
  • Cycle detector working ✓
  • FFT meter analyzer working ✓

Key Point: "Every algorithm outputs real numerical results, no simulation"

7. Advanced Features (30 seconds)

Quick mention:

  • Rete rule engine enforces constraints
  • Multi-objective scoring (meter + theme + novelty)
  • Online learning (TD updates during generation)
  • File upload for custom poetry styles

Closing Statement

"AG-TUNE demonstrates that poetry generation doesn't require massive neural networks. With classical algorithms - PCA, FFT, CYK parsing, reinforcement learning - we can create structured, metrically sound poetry from first principles."

Total Time: ~7 minutes

Backup Answers (If Asked)

Q: Is this using GPT or transformers? A: No. Zero neural networks. Pure algorithmic generation with Kernel PCA, TD learning, and beam search.

Q: How does training work? A: It builds co-occurrence embeddings from corpus, transforms them via polynomial kernel PCA into emotional space, then trains a TD-lambda value estimator to predict aesthetic reward.

Q: Can it scale? A: Yes - add more corpus via file upload. The algorithms are O(n³) for CYK, O(n log n) for FFT, both manageable for poetry-length inputs.

Q: What's the novelty? A: Multi-objective poetry generation combining 7 different algorithms (KPCA, TD, Rete, CYK, FFT, Floyd, Beam Search) in a unified framework without neural networks.

Technical Details Ready

  • Training process: See TRAINING_EXPLAINED.md
  • Source code: All in src/App.jsx (1,350 lines)
  • Algorithms documented inline with citations to theory

Good luck with the presentation! 🎯