← ALL WORK

GENERATIVE AI PRODUCT

Incredible Me.

Personalized children's books generated with AI

MY ROLEArchitecture and end-to-end engineering at Soluntech
STATUSIn production
STACKGemini 3.1 Flash Image · Python · Flask · React · OpenCV · SQLite
FROM THE SOURCE

A closer look at Incredible Me.

02 ORIGINAL VISUALS

01 / Review built into the bookIncredible Me production · demo account

Every page keeps its prompt and trace. The quality check flags pages where the face may still resemble the template, so a person reviews them before printing.

02 / Same scene, different childrenIncredible Me production · demo account

One story page, personalized for different children. The scene, style and secondary characters stay fixed while the main child changes.

INCREDIBLE ME / 01 OF 02
Incredible Me book view: a personalized cover and pages, each with prompt, trace, regenerate and refine controls, under a notice that two pages need review because the face may still resemble the template.

Review built into the book

Every page keeps its prompt and trace. The quality check flags pages where the face may still resemble the template, so a person reviews them before printing.

SYSTEM MODELSELECT A COMPONENT TO EXPLORE →
SELECTED COMPONENTParent photos

Uploads that define who the main character is.

01 / THE PROBLEM

Same child, every page, from a probabilistic model.

A personalized book only works if the child looks like the same child on every page. Image models are probabilistic: a face can drift back to the template, a mask can come back empty, and a page can fail without saying so.

02 / THE APPROACH

A reference first, then every page against it.

Parents upload photos, the system generates character references and the parent chooses one. Every page is then personalized with Gemini 3.1 Flash Image against that reference, double-page spreads are refined with both pages visible to the model, and a cover and a print-ready PDF close the book. An editorial manager controls books, styles, characters and locations, and every generation keeps its prompt and trace.

03 / THE DECISION

Plan for the model to fail.

I designed for the model to fail instead of assuming it would succeed. When model segmentation fails, a fallback chain locates the child, cuts a mask with OpenCV GrabCut and widens it if it comes back clipped. A local quality check reviews each page and retries automatically, using the failed image as a negative reference. Whatever still looks wrong is flagged for human review instead of being shipped.

04 / THE OUTCOME

In production, and honest about its doubts.

In production with users, built at Soluntech. Generation stays probabilistic, so the product surfaces its own doubtful pages, like the two flagged for review in the screenshot, rather than claiming perfect output.

SOURCES

The Incredible Me architecture documentation and the production demo account. Code and internal links stay private. This page summarizes what those documents show.

NEXT CASE STUDYForgeAI ↗