Content Automation · Solo-built

Book-Publishing Automation System

I designed and built a deterministic content pipeline that turns structured data into print-ready books. Verified data pools feed per-type generators whose output is correct by construction; a validation layer machine-checks every artifact; content is rendered through HTML/CSS to 6×9 print PDFs with automated preflight for page count, bleed and margins. The architecture already ships real products — a published puzzle-book series on Amazon KDP, and a Norwegian language-exercise engine built on the same core.

Python Deterministic generation Machine validation CSV data pipelines HTML / CSS templating Headless Chrome → PDF pypdf preflight Multi-agent workflow + QA gates Amazon KDP

The engineering is the project; the published books are proof the system ships real, print-quality products. The architecture is documented here — the proprietary data, prompts and market specifics are not.

5deterministic generators

100%outputs machine-verified

162 ppauto-preflighted 6×9 PDF

KDPpublished & shipping

The pipeline

Every book is produced by the same six-stage flow. Nothing is hand-typed into the book: data goes in one end, a validated, print-ready PDF comes out the other.

1 · Source

Verified data pools

Structured CSV (vocabulary, verb/noun tables, grammar rules), signed off once by a human authority.

→

2 · Generate

Deterministic generators

Per-content-type engines produce puzzles/exercises whose answers are computed by rule.

→

3 · Validate

Machine verification

Every output is reconstructed and checked; anything not provably correct is rejected.

→

4 · Render

HTML / CSS templating

Validated content is composed into a paginated book layout with a consistent design system.

→

5 · Export

Headless Chrome → PDF

Rendered at true trim size (6×9 + bleed) to a print-grade PDF.

→

6 · Preflight

Automated QA

pypdf checks page count, dimensions and no clipped/blank pages before it can ship.

How it stays correct

The core engineering idea: verify the data once, and generation is correct by construction. Correctness lives in the source tables, not in each generated page — so a human certifies a small amount of data, and the machine safely produces unlimited exercises from it.

Deterministic generation

Generators don't guess. Each answer is computed by rule from verified data — a conjugation from a verb table, an agreement from a gender, a crossword answer from its placement.

Validation layer

Every artifact is machine-checked before use: each crossword reconstructs, each hidden word is findable along its path, each answer key round-trips. Un-provable output is never emitted.

Rendering → print PDF

A single design system renders all content to a paginated 6×9 layout, exported via headless Chrome and auto-preflighted (page count = section count, no clipping, no blanks) so it passes print validation on the first upload.

Multi-agent workflow + QA gates

Bulk data drafting and precision engine code are split across tools, with automated QA gates between them (schema checks, de-duplication, "target-in-sentence" and value-range validators) that catch upstream mistakes before they reach a book.

What the system produces

Two independent product lines run on the same architecture — different content types, one engine.

Puzzle books — crosswords, word-hunts and matching, generated and validated.

Word-hunt grids — every hidden word machine-verified as findable.

Language exercises — conjugation, declension, agreement, word order.

Proof it ships: the puzzle-book series is published on Amazon KDP — the same pipeline that validated the content produced the interior that passed print preflight.

Browse the published series →