AI Publishing Platform

2026 – ongoing · AI layer, end to end · Production AI

Python, FastAPI, Pydantic, PostgreSQL, Azure, React, TypeScript, LLM tool-calling, Headless browser rendering, CI/CD

I built the AI layer of a multi-tenant publishing platform that turns loose copy and image sets into print-ready layouts. Vision models tag, rank, and place images using per-brand scores. Plain-language brand guidelines get compiled into validated patches, and the media library is searchable by embedding.

The layout itself stays deterministic. A model can propose a change; it only lands after schema and integrity checks, and every measurement is arithmetic.

Problem

A print layout that is nearly right still needs a person to fix it. So a model can suggest, but it cannot be the thing that decides where anything goes. Plausible is not good enough when the page has to fit.

Approach

  1. Layout selection and copy fitting are a deterministic search under hard constraints. Models propose; the search decides.
  2. Server and browser share one text-measurement implementation, so the preview and the print output agree by construction.
  3. One canonical document format feeds preview, export, hand-off to design tools, in-app editing, and the assistant. A new output target reads that format and needs no renderer changes.
  4. The chat editor runs a fixed set of editing operations on the server. Each is checked against the document before it applies, persistent changes ask for confirmation, and all of it can be undone.
  5. Jobs snapshot the original request, so a retry replays what the user asked for rather than half-mutated state. The runtime caps concurrency, recovers stuck jobs, and promotes artifacts tested in dev to production.

Results

  • In production, multi-tenant, with generated layouts going to real print.
  • New output targets ship without touching the renderer, because everything reads the one document format.

Limits

  • Client work under NDA. No product names, customers, scale, code, or screenshots here, on purpose.

Context

production platform
Multi-tenant
details generalised
Under NDA

Architecture

  1. Copy and images in
  2. Model layer proposes
  3. Validation + deterministic layout
  4. Canonical document format
  5. Preview, export, assisted editing