A client came to us with a specific problem: children learning languages need real-time pronunciation feedback, but human tutors can't scale. They wanted a platform where kids practice speaking, get instant AI-powered evaluation, and where parents and teachers can track progress — all in a web browser, no app downloads.
Six months and a three-person team later, Meraki was live. Here's what we built, the technical decisions we made, and what we learned along the way.
The challenge
On paper, the brief was straightforward: an interactive learning platform with voice exercises. In practice, the technical challenges were anything but.
The core problem was pronunciation scoring. The platform needed to capture a child's voice, send it to an AI model, evaluate accuracy, and return feedback — all in under two seconds, or the child loses interest. And this had to work across devices: laptops, tablets, phones, different browsers, different microphone qualities.
On top of that, three different user types needed three different experiences. Students needed gamified lessons. Parents needed progress dashboards. Teachers needed classroom management and the ability to listen to recorded attempts.
And the whole thing had to feel fun. A language learning tool for kids that feels like homework is dead on arrival.
The architecture decision
We split the system into three layers, each doing what it's best at:
Laravel backend — handles business logic, user management, lesson structure, progress tracking, and the admin panel. Laravel's ecosystem gave us authentication, queue management, and admin tooling out of the box, so we could focus engineering time on the hard problem: the AI layer.
React frontend — handles the interactive learning experience. Voice capture, real-time feedback display, gamification animations, and the parent/teacher dashboards.
Python microservice (FastAPI) — handles pronunciation evaluation. Takes an audio file, runs it through the speech model, and returns a score with specific feedback. This runs as a separate service connected via API.
This three-layer approach keeps each layer focused and replaceable. If a better speech model appears next year, we swap the Python service without touching the rest.
The hardest technical problems
Audio capture across browsers
Getting consistent audio recording across Chrome, Safari, Firefox, and mobile browsers sounds simple until you try it. Each browser handles microphone permissions differently. Safari on iOS has specific constraints. Some mobile browsers don't support certain audio codecs.
We spent the first two weeks just getting reliable audio capture working across all target devices. Not glamorous work, but foundational — everything downstream depended on clean audio input.
Pronunciation scoring with noise and accents
Children don't record in sound studios. They're in classrooms, living rooms, cars. Background noise, varying microphone distances, and accent variation all affect audio quality.
We built a pre-processing pipeline: noise reduction, volume normalization, and silence trimming. Without this, the model's accuracy dropped significantly in real-world conditions. The lesson: AI features need to be evaluated in real conditions, not lab conditions.
Keeping it fast enough for kids
A two-second delay is the maximum before engagement drops. Our initial implementation took 4-5 seconds. We got it under two through audio compression before upload, async processing with WebSocket feedback, and aggressive caching of lesson content.
What we delivered
For students: Structured lessons, voice exercises with real-time scoring, and gamified progress tracking. Each lesson is listen, repeat, score, celebrate.
For parents: Dashboards showing accuracy trends, areas needing practice, and configurable milestone notifications.
For teachers: Classroom management, student comparisons, audio replay of attempts, and homework assignments.
For admins: Content management, user management, analytics, and system configuration.
The AI layer: Pronunciation scoring returning both a numeric score and specific feedback on which sounds need work.
What we learned
Set measurable goals for AI features early. "Good pronunciation scoring" is vague. "Under 2 seconds with 85%+ accuracy on clean audio" is testable. Define thresholds in scoping, test against them throughout.
Validate the hardest technical risk first. We tackled audio capture in sprint 1 because everything else depended on it. If we'd discovered browser audio was unreliable in month 4, the project would have been in serious trouble.
Test with real kids, not just the team. First round of testing with actual children revealed issues we never anticipated: kids hold the microphone too close, they whisper, they giggle mid-word. User testing with your actual audience is irreplaceable.
Separate the AI from the platform. Running the speech model as a standalone service meant the platform stayed responsive even under AI load. We shipped three model improvements during development without redeploying the main application.
Design for observability from day one. The AI service, audio pipeline, queue system, and main application all had independent monitoring. When something went wrong, we pinpointed the layer immediately.
The result
Meraki launched on time, within budget, with a three-person team. The platform handles voice exercises, real-time AI evaluation, parent dashboards, teacher tools, and automated deployment — all on infrastructure that scales with usage.
For the full project details — technologies, features, and process — read the complete Meraki case study.
Could we build something like this for you?
The pattern here — Laravel platform + Python AI service + React frontend — applies to many domains beyond education. We've written about how long projects take, how to write the brief, and what to ask before choosing an agency.
Have a product idea that involves AI or real-time feedback? Let's explore it →