Project summary

I am building ManimAgent, an open-source multi-agent engine designed to produce an exhaustive, CC-BY library of ~25,000 K-12 Videos.

By pairing code-specialized LLMs with Vision-Language Models (VLMs) in a continuous Renderer-in-the-Loop architecture, we automate the generation, visual debugging, and cloud rendering of Grant Sanderson's Manim engine.

The central thesis of this project is that technical learning fails not from lack of practice, but from lack of spatial-geometric mental models. A student can memorize formulas for years without ever understanding why a derivative represents a tangent slope, why matrix multiplication rotates space, or how conservation of momentum functions geometrically.

Programmatic animation is the gold standard for building these intuitions, but human creation is bottlenecked at ~1–2 videos per month per animator.

With deterministic Python code acting as an exact, low-dimensional latent space for technical video, an automated pipeline can scale visual output across the entire K-12 STEM curriculum without sacrificing mathematical precision.

What are this project's goals? How will you achieve them?

• Exhaustive Curriculum Graph Mapping (~25,000 Modules): Taxonomize every micro-concept across K-12: Class → Subject → Chapter → Topic

• K12-ManimBench: Build an open-source benchmark of 1,000 foundational conceptual bottlenecks to stress-test agentic code generation.

• Agentic Renderer-in-the-Loop Architecture (ManimAgent): Build a production-grade pipeline where LLMs write Manim Python scripts, compile them in headless Docker containers, and use VLMs to visually inspect rendered frames for text overlap, off-screen clipping, or camera pacing bugs.

• Public 25,000-Video Repository: Host and publish all modular, under a Creative Commons (CC-BY) license.

How will this funding be used?

• LLM & VLM Inference API Costs ($7,500): Funding multi-turn agentic reflection loops for script generation, visual frame inspection, and automated bug-fixing.

• Human-in-the-Loop Validation Bounties ($2,500): Targeted bounties for K-12 teachers and undergraduates students to spot-check generated videos for subtle pedagogical edge cases or visual ambiguities.

Who is on your team? What's your track record on similar projects?

Tamzid R (Project Lead): Co-founder, Borno AI (AI Tutor for Bangladeshi Curriculum) & Testly (Largest SAT/ACT Bank)

• manim-ast-parser: I built and open-sourced an Abstract Syntax Tree (AST) parser for Manim that extracts variable bounds, animation durations, and visual bounding boxes directly from Python ASTs to prevent text collision and overlap during LLM script generation.

• Technical Research on Latent Code Spaces: I authored a widely cited technical writeup outlining why code execution acts as a strictly superior latent space for technical media generation compared to frame-by-frame pixel diffusion models.

• Benchmarking & Optimization: I benchmarked frontier code LLMs on Manim generation tasks, demonstrating that adding automated visual feedback loops (Renderer-in-the-Loop) drops execution and visual layout error rates from 65% down to <8%.

• Previously built AI Tutor in NCTB Curriculum (https://borno.ai) and SAT/ACT Question Bank (https://www.testly.fyi)

What are the most likely causes and outcomes if this project fails?

Pixel-based video diffusion models (e.g., Sora, Runway, Veo) are the wrong paradigm for educational and technical video generation.

Industry capital is overwhelmingly flowing into generative pixel models. While these models excel at cinematic realism, they are fundamentally ill-suited for STEM education because they lack exact symbolic conservation laws and continuous geometric invariants. In a diffusion video explaining calculus or physics, coordinate axes warp unpredictably, numbers hallucinate sub-pixel artifacts, and geometric transformations lose precise linearity.

Programmatic rendering (text-to-code-to-render via Manim or SVG DSLs) is the correct abstraction. Code acts as a deterministic, exact representation of truth. An LLM operating on code guarantees ground-truth logic—a triangle remains geometrically rigid, axes scale deterministically.

Instead of spending billions training 100B+ parameter generative diffusion models to guess pixels, we should use lightweight code models as symbolic programmers driving deterministic rendering engines.

How much money have you raised in the last 12 months, and from where?

$25,000 in Google Cloud (GCloud) Compute Credits (Secured to run distributed Docker rendering nodes for mass video compilation).