Case studiesEliacover

Case study · Elia Interactive’s own product

Eliacover: a complete AI product, conceived, designed, built and run by Elia Interactive.

AI-personalized books where each person’s real face stays the same on every page. Digital, print or audio, in 6 languages, with payment, printing and delivery. An idea born at Elia Interactive and carried through from start to finish: it is also where the way of working shows.

Client products stay confidential. Eliacover belongs to Elia Interactive, so it can be shown in full, from the AI engine to cost tracking.

Visit eliacover.com (new tab) For technical teams
Example book · fictional charactersCover
The user gives
Photo provided: MomPhoto provided: LouPhoto provided: Lou drawing

“Lou is afraid of the dark. One night, Pompon, her cuddly toy, starts to talk…”

They receive
The same face on every page
Listen to page 2Audiobook · multiple voices
The showcase example book: three photos and an idea in, an illustrated, narrated book out. Fictional characters.
01The problem

A personalized book only works if it’s the same person on every page.

Ask an AI tool for ten images of the same person: you get ten slightly different people. For a book, that’s a deal-breaker.

Constraint 1

The same face, on every page

Recognizable from the cover to the last page, whatever the setting or the scene.

Constraint 2

Every book is one of a kind

The story, the illustrations and the layout are produced for a single order, from each person’s photos and idea.

Constraint 3

Several formats and languages

Digital book, printed book delivered to the door, audiobook. 6 languages to choose from for each book.

A typical image tool

The face changes
The starting photoThe photo
Page 1
Page 2
Page 3
Page 4
Schematic illustrationFrom one page to the next, it’s not really Lou anymore.

Eliacover

The same face
The same starting photoThe photo
Page 2
Page 9
Page 10
Page 23
Left, what a typical image tool produces (schematic). Right, the real pages of the example book.
02The solution

Photos and an idea in. A checked book out.

Six AI steps chained between the photos and the book, each one checked. An engine in production, with no servers to manage, on AWS. In plain language, here’s what happens.

The user gives
  • A few photosof the people in the book
  • An ideain a few words
  • Or documents, a voice noteto tell it their way
The AI engine6 steps
  1. Recognize the facesin the photos, only with the people’s consent
  2. Write the storyfrom the idea, in the chosen language
  3. Illustrate every pagewith the same face, page after page
  4. Check every imagelikeness, errors, page redone if needed
  5. Give it a voicefor the audiobook, multiple voices
  6. Lay it outready to read, listen to or print
They receive
  • A digital bookto read on any screen
  • A printed bookprinted and delivered to the door
  • An audiobooknarrated, page by page
  • An animated coverthe cover, in motion
The engine, from input to output. The AI model names are in the technical section.
Cover: “Lou et la lanterne attrape-ombres” (Lou and the Shadow-Catching Lantern)Cover
Page 2: Lou draws rabbits with her momPage 2
Page 9: Dad brings the lanternPage 9
Page 10: Lou lights the lanternPage 10
Page 13: Lou and Pompon above the rooftopsPage 13
Page 14 of the example bookPage 14
Page 23: Lou falls asleep by the lanternPage 23
Page 24: good night, LouPage 24
Eight pages of the example book: Lou stays Lou, from the cover to the last page. Fictional characters.
Another starting point · the photo album

Hundreds of travel photos. The engine understands the trip.

The photos arrive all mixed up. The engine works out where and when they were taken, understands what they show, and tells the story, chapter by chapter.

  1. Dates and places

    Read from the photos themselves, then turned into real place names. The route takes shape on a map.

  2. What each photo shows

    A vision AI describes each image. Duplicates and blurry or failed shots are set aside.

  3. Faces, with consent

    Only if the user agrees: no faces cut off, and photos grouped by person.

  4. The story and the layout

    Chapters, highlights, a lead photo for each one. The engine only asks for what it can’t guess.

GPS location and file names are never sent to the AI models.

A travel album being analyzed. Illustrative example.
03In your image
In development

Documents and a photo. A book where you find yourself on every page.

A guide, recipes, memoirs, the story of a company. The user uploads their documents, voice notes and a photo of themselves. The engine reads everything, then puts them in the scene on every page.

  1. Every document, read in full

    Hundreds of pages, even scanned or handwritten, plus voice notes. Nothing is left out.

  2. Split into passages, in a vector index

    Each document is split into passages that keep their source and page. They are stored in an Amazon S3 Vectors index of their own, one per book.

  3. Written from the real words

    For each page, the engine fetches the passages that talk about it. Facts and quotes are checked against the documents, not made up.

  4. The person, on every page

    The same face from one page to the next, in their world.

A secure index, one per book

Each book has its own private, isolated index: a book can never read another one’s documents. The index is deleted along with the files.

A page written from the documents, each passage linked to its source. Illustrative example, fictional character.
04Quality control

A second AI checks every page before it is shown.

An AI that creates images from photos has to be kept in check. Every page goes through automated checks; whatever fails is redone, fixed or flagged.

A page of the example book being checked
Before showing a page

What the engine checks

  1. The face is compared with the original photos. Too far off, and the image is redone.
  2. Several candidate images are measured, and the best one is kept.
  3. A failed image is regenerated, then the face gets targeted fixes.
  4. A second AI reviews the page: bad compositing, stray or duplicated text, pasted-on look.
  5. No one is missing, no one is extra.
RegeneratedFixedFlagged
A page of the example book goes through the checks, one by one.
Data

Photos, protected

  • Private, encrypted storage
  • Temporary access links, limited to a single book
  • Face analysis only with explicit consent
  • Uploaded photos go through an automatic filter, then a moderation queue
  • Depending on the consent given, AI providers commit not to train their models on the photos
  • Data can be exported, accounts deleted on request, photos deleted automatically
Code

Built with AI, reviewed and tested

≈19,700

automated test cases

118

behavior scenarios

  • The AI writes the code and the tests, a human reviews and decides
  • End-to-end tests, just like a real customer
  • A security audit
Releases

Nothing ships without sign-off

  • Separate environments: changes are tried first, customers come after
  • Human sign-off required before every release
  • Dated versions and a planned rollback
  • If the AI refuses an image, the engine falls back step by step and the page is flagged
05Monitoring

Sell, print, deliver, and track what it costs.

A product doesn’t stop at the AI engine. Selling a real book took everything else too, designed and delivered with the same care, and monitored once live.

Admin console · exampleIn production
Printed bookOrder · 24 pagesPrinted
AudiobookOrder · multiple voicesReady
Uploaded photoModeration queueTo review
Page 14Flagged by the engineRedone
Approved Printed Shipped Delivered
Production cost of one book68% of budget
Budget per book and per action · alert before overrun
The console: orders, moderation, delivery and costs on a single screen. Illustrative example, fictional data.
6languages

to choose from for each book, digital, print or audio.

Payment
Secure online payment, at checkout.
Printing
From the print-ready file to the book delivered to the door.
Costs
A budget per book and per action, with an alert.
Operations
Orders, photo moderation, flagged pages and costs, from a single console.
Legal
Trademark registered with INPI in France, legal notice, GDPR and consent.
Brand
Name, logo, design system and product videos.
06For your product

What the same method can do for your product.

Eliacover went through every step of the path offered to clients. You enter at the step that suits you; each one ends with a deliverable and a decision: continue, adjust or stop there.

Stage 1

Propose and prototype

On EliacoverThe purchase journey, the architecture, the choice of AI models, what each book costs to produce.
  • A specification reviewed with you
  • AWS architecture and choice of AI models
  • Delivery plan and estimate
Propose and prototype
Stage 2

Build and go live

On EliacoverA multi-step AI engine, each step checked, automated tests, separate environments.
  • Product live on your AWS account
  • Automated tests and AI quality checks
  • Code, documentation and access handed over
Build and go live
Stage 3

Monitor and improve

On EliacoverCosts tracked per book and per action, alerts, one console to run it all.
  • A diagnosis with numbers, backed by measurements
  • Recommendations ranked by impact
  • The most useful ones put in place, then monitored
Monitor and improve
The code is yours, and your product runs on your own AWS account. Services in detail
07Technical teams

The engine, the infrastructure and software quality.

For technical teamsThe services and models used, in one minute.

The AI engine

Text and story
Amazon Bedrock (Claude) and Gemini models
Same-face illustrations
Gemini image models
Face analysis and grouping
Amazon Rekognition
Audiobook voices
ElevenLabs
Voice notes
Amazon Transcribe
Reading the photos in an album
Gemini vision models
Places and routes
Amazon Location Service
Search across the documents
Amazon S3 Vectors and Amazon Titan Embeddings
Animated covers
Google Veo

The infrastructure

  • Fully serverless on AWS, Paris region for storage
  • The engine orchestrated step by step
  • Infrastructure defined as code
  • Separate environments, test then production
AWS LambdaAWS Step FunctionsAmazon DynamoDBAmazon CloudFrontAWS CDK

Software quality

  • ≈ 19,700 automated test cases
  • 118 behavior scenarios
  • End-to-end tests
  • Human sign-off before every release, dated versions, rollback
  • Security audit
  • Continuous integration

Names only: no configuration, no costs, no instructions given to the models.

Your product

Tell me about your product. I reply within 2 business days.

Let’s talk about your product