Trusted Tech Leader/ Case studies

Case study · AI & automation

AI textbook ingestion and study reader

Turns scanned textbook chapters into a searchable reader with AI answer cards, glossary links, figures, quizzes and flashcards.

Sector
Education technology
Category
AI & automation
Maturity
In active development with a tested ingestion pipeline; demo-content mode and a local database mode

At a glance

Who it's for
Secondary-school students studying from a prescribed textbook, and the content reviewer preparing it.
Core stack
Next.js, PostgreSQL, Gemini models, Zod
What sets it apart
A review-aware pipeline that converts a scanned textbook into linked, interactive study material.
On this page

What it looks like in use

A student studying one chapter: practice questions, answers, key words and the four ways to study.

Class 11 Biology: chapter study
Class 11 BiologyRead Learn Quick revise Test me
Notice shown above the questionsThese answers were written by AI from the textbook text and have not been checked by a teacher yet. Use them to practise, and confirm important facts in the book.
Question cardWhat is a cell?
Show answerOpen
The cell is the basic unit of life. Every living thing is made of one or more cells.AI-written answer
Exam pointsTap to open
Revise laterGot it
Chapter progress
Marked Got itMost cards
Marked Revise laterA few cards
Not tried yetSome cards
Key words
CellHear it, see the meaning
OrganelleHear it, see the meaning
MembraneHear it, see the meaning
Figure from the textbook
Simple meaningAI-written
Illustrative screen sample data, not client data

Real screens from the working product (sample data)

A study card ("Define biodiversity") in Learn mode, with a tapped key word opening a panel showing its pronunciation, a plain meaning and a translation marked as a draft.
Real screen, sample data The study card in Learn mode: tapping a key word opens its pronunciation, a plain meaning and a translation, flagged as a draft to check with a teacher. Demonstration text. Photo: lubasi / Flickr, CC BY-SA 2.0.

A study session, step by step

  1. Sara opens her biology chapter

    She picks a way to study: read the text, learn with cards, revise quickly or test herself.

  2. She taps a word she does not know

    A short, plain meaning opens beside the text, so she does not have to leave the page.

  3. Each answer says who wrote it

    AI-written answers carry a clear label, so she knows to check important facts in her book.

  4. She photographs a page from her book

    The page lens reads the photo and gives her an AI analysis of the page.

  5. Her progress is kept for next time

    Cards she marked Got it and Revise later are remembered, so she can pick up where she stopped on her device.

The challenge

Students study from scanned PDF textbooks that are hard to search. There is no linked glossary, no practice questions, and figures are locked inside page images.

Turning a book into structured study material by hand is slow. AI-generated study content is also risky, so provenance and review are treated as first-class concerns, with a warning about copyright on source scans.

The solution

An ingestion pipeline renders PDF pages, extracts text into chapters, topics, sections and paragraphs, and asks AI models to generate answer cards, glossary terms and figure bounding boxes.

Results are validated against schemas, cached, and imported into the database in all-or-nothing steps that are safe to repeat, with a trial-run mode. A Next.js reader then offers inline glossary terms and figures, quizzes, flashcards, bookmarks, progress, search and a page-lens tool that analyzes a photographed page.

For your technical teamhow it works under the hood

Ingestion stages across model, cache and database

Ingestion stages across model, cache and databaseSwimlane diagram with three lanes: Gemini model, cached files, and PostgreSQL, across six ordered stages. Stage 1 loads chapter text into PostgreSQL. Stage 2 has the model write answer cards, saved to the database. Stage 3 has the model write glossary terms into a terms file without touching the database. Stage 4 reads that file back and writes terms and paragraph markers in one transaction. Stage 5 has the model locate figures and writes cropped figures and a figures file. Stage 6 reads the figures file back and places figures in sections, linked to terms.Gemini modelAI stepCached filesreviewable, no databasePostgreSQLchapter content1 Text2 Cards3 Terms4 Link terms5 Figures6 Link figuresChapter texttext structureAnswer cardsmodel writesQuestion cardssaved to databaseGlossary termsmodel writesterms.jsonno databaseterms.jsonread backTerms, markersone transactionFigure boxesmodel locatesCropped figuresfigures.jsonfigures.jsonread backSection figureslinked to terms
Stages run left to right. For terms and figures, the model's output is cached to a file first and a separate import step writes it to the database. Answer cards are saved directly.

Key features

What the system does, as built in the codebase.

  • 01

    Chapter reader

    A read view with sections, paragraphs and figures placed between the text.

  • 02

    AI answer cards

    Generated question cards with summaries, points and multiple-choice explanations.

  • 03

    Auto-linked glossary

    Whole-word term markers are inserted into paragraphs and linked to cards and diagrams.

  • 04

    Figure extraction

    A vision model locates figures on rendered pages and crops them to images.

  • 05

    Quizzes and flashcards

    Practice tools with local progress tracking.

  • 06

    Full-text search

    Search over paragraphs using a search index built for the purpose.

  • 07

    Page lens

    Upload a photo of a page for AI analysis, with rate limiting.

  • 08

    Word lookup and speech

    Vocabulary support with word lookup and speech.

How it fits together

From a scanned PDF to a reviewed, searchable reader.

  1. Textbook PDFpage rendering
  2. AI ingestionanswer cards, terms, figures
  3. Validation & cacheZod schemas
  4. PostgreSQLchapters, cards, glossary
  5. Next.js readerread, quiz, search
  6. Page lensphoto analysis API
Pages are rendered and passed to AI ingestion, which produces cards, terms and figures. Output is validated and cached, then stored in PostgreSQL. The reader serves students from that data; the page-lens API is a separate route that sends a student's photo to the model.
  • Ingestion scripts write through idempotent import transactions.
  • The reader reads through a repository that can fall back to demo content.

Implementation notes

Engineering decisions worth knowing about.

  1. Provenance and review states

    Content carries explicit provenance (demo, approved, AI-generated) and review status (draft, needs review, approved).

  2. Repository pattern

    In-memory demo content and a PostgreSQL implementation are chosen by environment.

  3. Database security

    A row-level security migration is paired with server-only database access.

  4. Staged, safe ingestion

    Cached stages use schema-validated files, partial-run guards, transactions and dry-run.

  5. Model fallback

    Multiple models are tried when figure detection fails.

  6. Image safety

    Image features use safe fetching, license checks and request limits.

  7. Tested

    Extensive unit tests plus database integration tests.

Technology stack

Taken from the project's dependency and config files.

Frontend
  • Next.js
  • React
  • Tailwind CSS
  • shadcn/ui
  • Base UI
Backend
  • Next.js route handlers
  • Node.js
  • TypeScript
Data
  • PostgreSQL (pg)
  • Full-text search
  • Row-level security
AI & PDF
  • Gemini models
  • mupdf
  • pdf-lib
  • sharp
Validation & testing
  • Zod
  • Vitest
Infrastructure
  • Docker Compose for local database
For your technical teamtechnology layers

Technology layers, reader to infrastructure

Technology layers, reader to infrastructureLayered diagram of the stack. Frontend: Next.js, React, Tailwind CSS, shadcn/ui, Base UI. Backend: Next.js route handlers, Node.js, TypeScript. Data: PostgreSQL, full-text search, row-level security. AI and PDF: Gemini models, mupdf, pdf-lib, sharp. Validation and testing: Zod and Vitest. Infrastructure: Docker Compose for the local database.Frontendthe study readerNext.jsReactTailwind CSSshadcn/uiBase UIBackendroute handlersNext.js route handlersNode.jsTypeScriptDatacontent and searchPostgreSQL (pg)Full-text searchRow-level securityAI & PDFingestion scriptsGemini modelsmupdfpdf-libsharpValidation & testingchecks on every stageZodVitestInfrastructurelocal setupDocker Compose for local database
The reader and its route handlers sit on PostgreSQL. The ingestion tools fill that database offline, and validation and testing wrap every layer.

Business value

What the system enables. No numbers are claimed.

  • Converts static scanned material into interactive, searchable study content.
  • Keeps AI output reviewable through explicit provenance and review states.
  • Makes chapter ingestion repeatable and safe to re-run.
  • Works without a database through demo mode, lowering the barrier to try it.

Need a system like this built?

Tell us what it needs to do and we will scope it with you on a call.

Discuss a build