Project

My Story Book Video

A generative video product that turns a child's details into a personalised animated storybook — built on an LLM + video-model pipeline, shipped behind an evals harness

What It Is

My Story Book Video generates personalised story videos for children. A parent enters a few details — name, age, interests, the kind of adventure they’d love — and the system produces a complete, narrated, animated storybook video starring that child.

Every part of the output is generated: the story itself, the scene-by-scene visuals, and the narration. Nothing is templated — which is exactly what makes it hard to ship.

The Pipeline

The product is a multi-stage generative pipeline:

  • Story generation — an LLM writes the story from the child’s details, constrained to a structured scene-by-scene output (setting, characters, action, narration text per scene) rather than free prose.
  • Visual generation — each scene is rendered with generative video models, with prompt engineering to hold character appearance consistent from scene to scene.
  • Assembly — narration, timing, and scenes are stitched into the final video.

The Evals Harness

The interesting product problem: generative output can’t ship on vibes. A story that’s charming nine times out of ten still fails a parent the tenth time. So the pipeline runs behind an evals harness:

  • Rubric-based LLM judges score each generated story on narrative quality, age-appropriateness, and personalisation — did the child’s details actually shape the story, or just get name-dropped?
  • Deterministic graders enforce the structural contract: scene counts, output schema, duration bounds, and content-safety checks.
  • Ship gates — a pipeline change only rolls forward when eval scores clear thresholds against a fixed test set of inputs.

This is the same discipline as A/B testing a growth funnel, applied to non-deterministic systems: define what “good” means, measure it before users do, and gate releases on it.

Why I Built It

Partly for the product itself — and partly because building and evaluating LLM pipelines end-to-end is the core skill of AI product management right now. Writing rubrics, choosing thresholds, and trading generation cost against quality are product decisions, not engineering ones.