Arlos is an AI-powered education platform built for schools, giving every student a personal AI tutor and teachers a dashboard that automates the admin work. Our challenge was helping Arlos generate a library of custom, on-brand lesson visuals fast enough to hit a hard launch deadline, without adding headcount to a small founding team.
Arlos needed 100 to 140 interactive lesson intros per subject per year, each with custom artwork in the platform's own visual style. Producing that volume by hand wasn't realistic for a small team, and it stood between Arlos and its launch.
At the same time, Arlos wanted broader, more flexible access to model infrastructure for the AI tutor at the heart of the platform, one that needed to sound encouraging and age-appropriate for students aged 10 to 14, not just technically correct.
Migrating the tutor to AWS Bedrock
We evaluated Claude models on Bedrock against Arlos's own tutor and classifier prompts, testing for reasoning quality rather than just wording match. Claude Sonnet came out ahead, demonstrating correct intent-matching on real tutoring scenarios, and was selected to power the live tutor.
Teaching a model Arlos's visual style
We fine-tuned a Flux2 image model with LoRA on Fal.ai, training it on real lesson artwork supplied by Arlos so generated visuals match the platform's established look instead of generic AI art.
Building a quality gate, not a bottleneck
Manual art review doesn't scale to hundreds of lessons. We built an automated quality gate checking style similarity, color palette, transparency, and embedded text on every generated image, so only on-brand visuals move forward without a human reviewing each one.
The quality gate hit near-perfect accuracy on style-matching, with a 0% false-accept rate, screening generated lesson art automatically instead of relying on manual review. Arlos's tutor now runs on Claude Sonnet via Bedrock, and the LoRA model was trained and validated against real brand assets, giving the team a repeatable pipeline for lesson visuals as they head toward their September launch.
Tell us more about yourself and what you're got in mind.
Prefer email? hello@apexlab.io
Or even a quick call?
We don’t care about fancy CVs or long interview processes. Provide some basic information, let’s have a quick chat, and then show us what you love doing the most: write some code and create a product.
