A self-hosted AI textbook system that generates grade-appropriate, citation-verified chapters across any subject, using a hybrid of free local AI and paid models to keep costs under $35/month.
Textbooks are expensive, outdated before they're printed, and rarely customized for specific student populations. Community colleges and smaller institutions can't afford to commission custom materials, and their students can't afford $200 textbooks.
Open Educational Resources (OER) exist but are hit-or-miss in quality, rarely aligned to specific curricula, and take significant faculty time to curate. Meanwhile, AI can generate educational content, but without guardrails, it hallucinates citations, invents facts, and produces content at the wrong reading level.
The challenge isn’t generating text. It’s generating trustworthy educational content at the right level, with real citations, at a cost that makes sense.
Students spend hundreds per semester on textbooks. Institutions need affordable alternatives.
The same textbook serves elementary through graduate students. Reading levels don’t match learners.
Generic AI generates convincing but fabricated citations. Academic content needs verified sources.
Creating custom materials from scratch takes months. Curating OER content is nearly as slow.
Designed a cost-optimized routing system: free local Llama 3.1 (8B) handles 95% of tasks (objectives, outlines, summaries, keywords), while paid Gemini Flash handles only complex chapter content. Automatic fallback ensures reliability.
Built a structured pipeline: objectives → tone config → outline → source retrieval (RAG via Qdrant) → section-by-section generation → citation validation → summary → key terms → practice questions.
Every citation is validated against CrossRef and Semantic Scholar APIs. Fabricated references get flagged and removed. Only real, verifiable academic sources make it into the final content.
9 tone profiles (Academic, Conversational, Storytelling, Socratic, Humorous, etc.), 8 style modifiers, and 5 formality levels let faculty tailor content to their teaching style and student population.
Let's talk about how AI-generated, citation-verified content could work for your institution.
Start a ConversationGenerates content calibrated to 5 reading levels: Elementary (Lexile 200-800), Middle School (800-1000), High School (1000-1300), College (1300-1600), and Graduate (1400-1800).
Mathematics, Biology, Chemistry, Physics, Earth Science, English, History, Geography, Social Studies, Computer Science, Art, Music, Economics, Psychology, Philosophy.
From Academic (“Research has demonstrated...”) to Humorous (“Plants are running their own solar-powered restaurants!”) to Socratic (“What do you think happens when...”).
Every reference validated against CrossRef and Semantic Scholar. No hallucinated sources. Includes DOIs and publication metadata.
Each generated chapter includes objectives, section content, summary, key terms with definitions, and practice questions. Export to PDF, EPUB, HTML, or DOCX.
Intelligent LLM routing keeps monthly costs around $35. Real-time cost dashboard shows exactly how much each generation costs at the token level.
Configure subject, grade level, tone, and topic. Watch the 9-step pipeline progress in real-time.
Side-by-side view of the same concept explained in Academic vs. Conversational vs. Storytelling tone profiles.
Real-time token-level cost tracking showing the hybrid LLM routing efficiency.
We were spending thousands on textbook licenses every year. This system lets us generate custom chapters at the reading level our students actually need, with real citations. The Socratic tone profile has been a game-changer for our developmental courses.
The initial assumption was that you need a frontier model for everything. In practice, Llama 3.1 8B running locally handles outlines, summaries, key terms, and practice questions just as well. Only complex narrative generation benefits from the paid model.
Faculty who piloted the tool cared more about whether the content sounded like them than whether it was technically perfect. Adding the 9 tone profiles tripled pilot adoption. The content was always accurate, but it needed to feel right.
The first version generated plausible-sounding citations that were completely fabricated. Adding CrossRef/Semantic Scholar validation took a weekend of work but was the single feature that made faculty trust the output enough to actually use it.
Tell us about your textbook challenges and let's explore what AI-generated, citation-verified content could look like for your courses.
No pitch. No pressure. Just a conversation about what might work.