How laIA works: the pipeline, the models, and the privacy guarantees

The pipeline, step by step

When you upload photos and pay, the pipeline starts. Step one (~10 seconds): Amazon Rekognition scans each photo against child-safety policies. Inappropriate images never enter the system; the transaction is reversed and we notify you.

Step two (~30 seconds): OpenAI gpt-5 reads the mission you picked (or the free text you wrote), the photos as visual reference, and plans the narrative structure of the 20 pages. Your kid's face becomes an internal reference the pipeline reuses on every image generation — this is what keeps the protagonist visually consistent across the whole book.

Step three (~5 minutes): OpenAI gpt-image-2 generates the 20 illustrations, one per page. Quality is pinned to 'low' so a book doesn't take 30 minutes — an explicit time/resolution trade-off. Each illustration passes another moderation round before being accepted.

Step four (~10 seconds): ReportLab composes the 1-3 MB PDF with proper typography, print-ready bleeds and a cover. Step five (instantaneous): AWS SES sends the email with the download link. Total: 5-10 minutes from payment.

Which models we use and why

Text

OpenAI gpt-5. Strong reasoning for long narratives; keeps coherence across pages. Claude / Gemini work as alternatives; we stay with gpt-5 for API ergonomics.

Image

OpenAI gpt-image-2. Better character consistency across series than MidJourney or SDXL — critical for your kid to have the same face on page 4 and page 17.

Moderation

Amazon Rekognition. AWS-native (lives in the same region as the upload bucket — no network hops, no third parties).

PDF composition

ReportLab (Python). Open source, no external services. Produces lightweight PDFs (1-3 MB) that open on any device.

Infrastructure

AWS region eu-west-1 (Ireland) — kid data never leaves the EU. Step Functions orchestrates the pipeline; S3 stores photos with a 30-day lifecycle rule; SES sends the email.

Privacy guarantees — what we promise and why

Three guarantees that set this product apart from generative-AI alternatives that feed on user data:

We don't train models on your photos. OpenAI has a contractual no-training policy for API calls (not consumer ChatGPT) — it's in the DPA we signed with them. Amazon Rekognition doesn't use inference data for training either. Our code doesn't keep copies anywhere else. The photo arrives, gets processed, gets deleted.

Photos are auto-deleted after 30 days by architecture. It's not our policy — it's an S3 lifecycle rule painted on the bucket. Even an internal operator can't override it; objects disappear automatically. This covers source photos and intermediate renders.

We comply with GDPR, LSSI-CE, LOPDGDD and EU AI Act Art. 50. The product was built from day one for Europe with kid data — the strictest regulatory regimes. To delete your data before 30 days, email privacy@cuentoconlaIA.com.

How this differs from Gemini Storybook

Gemini Storybook launched in August 2025 and is free — that matters. But it does something different: writes a story with a generic mascot that isn't your kid. Here, your kid is the protagonist on every one of the 20 pages.

Other technical differences that matter in daily use: Gemini's output is a heavy file that slows the browser and doesn't print well; ours is a lightweight PDF designed for sharing and printing. Gemini loses character consistency in long format; we keep it because the pipeline reuses the face reference. Gemini is one prompt and a prayer; we are a guided wizard.

And the least visible but most important piece: Gemini can change its data-use policy overnight (it's a consumer product). Our API contract with OpenAI is stable and auditable.

Read also

To see the output before paying, open /example and look at Pol's book with his plushie Buli. To make yours, the wizard takes 30 seconds.

Frequently asked questions

Read also