Text-to-Animation Engine (TTA)

1. The Problem

Creating a single high-quality educational animation requires subject matter experts, structured scriptwriting, animation engineering, and voice + post-production — a largely sequential, human-intensive workflow that is impossible to scale.

2. The Insight: LLMs Lack Spatial Awareness

Early experiments asking LLMs to write Manim scripts directly failed ~60% of the time. Models called non-existent functions, placed text labels over diagrams, and produced invalid syntax. Pure generation is unreliable for structured animation code.

3. Architecture: Neuro-Symbolic Pipeline

The solution is a hybrid approach: use LLMs for high-level reasoning and content synthesis, but confine them within strict, deterministic code scaffolds.

• Subject Classifier — routes topic to domain-specific template engine (confidence > 85%)
• Template Engine — subject-specific Manim scaffolds for Maths, Physics, Chemistry, Organic Chemistry, CS
• Wikipedia Fallback Agent — grounded generation for low-confidence or unsupported topics
• Orchestration Layer — manages routing, validation, regeneration, rendering, and storage
• TTS + Sync Layer — audio narration with <100ms video synchronisation

4. Confidence-Gated Routing

def route_topic(topic: str, subject: str, confidence: float):
    if confidence >= 0.85:
        return use_domain_template(subject, topic)
    else:
        return wikipedia_fallback_pipeline(topic)

5. Template Architecture

def _get_template_for_subject(self, subject: str) -> str:
    template_map = {
        "mathematics":          MATH_TEMPLATE,
        "physics":              PHYSICS_TEMPLATE,
        "chemistry":            PHYSICAL_CHEMISTRY_TEMPLATE,
        "organic_chemistry":    ORGANIC_CHEMISTRY_TEMPLATE,
        "computer_science":     COMPUTER_SCIENCE_TEMPLATE,
    }
    return template_map.get(subject, MATH_TEMPLATE)

6. Results

• 99.4% success rate (up from ~40% with pure generation)
• P95 latency: 85 seconds for a fully rendered, narrated video
• Deployed across PCM subjects at Edza.ai

7. Key Learnings

• Neuro-symbolic is more robust than pure generation for structured outputs
• Confidence-gated routing prevents template mismatch failures
• Isolating failure boundaries enables independent component testing