Curriculum AI Pipeline
← Project Index Curriculum AI Pipeline

Curriculum AI Pipeline

Converts raw curriculum documents (DOCX, PPTX, PDF) into fully-specified, AI-powered interactive lessons through a multi-stage pipeline, where each stage produces a durable artifact that feeds the next.

Archived Started: 2025 Updated: Jan 2026

Overview

An internal work tool that takes the messy reality of curriculum -; Word docs, PowerPoints, PDFs, all formatted for human teachers -; and turns it into a fully specified, AI-driven interactive lesson. The design principle is that nothing happens in one giant prompt: the pipeline is a chain of stages, and each stage writes out a real, inspectable artifact that the next stage reads. You can stop, review, and resume anywhere along the line.

How It Works

The pipeline runs in Python and uses Claude for every generative step. It starts by getting documents into a form a model can actually read, then climbs from extraction to analysis to design to a finished lesson config:

  • Document processing (1-;4) -; DOCX/PPTX converted to PDF via Office automation, compressed with Ghostscript, then turned into companion text; images optionally described by a vision pass. Artifact: paired .pdf / .txt files.
  • Analyze (5) -; deep-reads a lesson and pulls objectives, standards alignment, vocabulary, concepts, and prerequisites. Artifact: an analysis folder of JSON + markdown.
  • Design (6) -; a two-pass step that first deterministically picks a story arc, cast, and game mechanics, then generates scenes, backstories, and a badge system. Artifact: LESSON_DESIGN.json + markdown, with review loops.
  • Generate config (7) -; transforms the design into a validated ES-module lesson.config.js: scene dialogue, AI prompts for the in-lesson tutor and teacher characters, engagement-check questions, and a HeyGen video manifest.
  • Build (8) -; the interactive HTML/JS lesson itself. Still in progress.

The design stage leans on three reference libraries -; a catalog of story arcs, one of characters, one of game mechanics -; so the "make this a game" choices are grounded in a curated set rather than invented fresh each time. Video is produced through a HeyGen avatar pipeline driven by the config's voice and avatar IDs.

Current Status

Archived as an internal prototype. Stages 1 through 7 are implemented and demonstrated end to end on a worked Algebra 1 example, with the interactive build stage (8) the main piece left unfinished. A companion tool gives an AI tutor progressive, drill-down access to curriculum context without dumping an entire scope-and-sequence into the prompt.

  • Pipeline runs from raw documents through to a validated lesson config.
  • Worked example: Algebra 1, with a full set of lesson folders.
  • Interactive HTML/JS lesson build is the remaining stage.