Two first-time founders came to us with real conviction: Arabic-speaking students in the Kingdom deserved an AI tutor built for them from the ground up, not an English product with a translation layer bolted on. We partnered with them at the earliest possible stage — shaping the product strategy, the feature set, the brand, and the design system alongside the full engineering build underneath.
The result is Mualim, a bilingual (Arabic/English) AI tutoring platform for Grades 4–9, spanning three connected applications — a Student learning experience, a Parent oversight dashboard, and an AI-powered Admin panel for curriculum and content control — with a curriculum-grounded AI tutor that teaches by text and real-time voice, in either language, autonomously generating the questions, images, and tests it needs as it goes.
We took the vision from first principles to a delivered MVP in under four months, running the full build — product, design, and engineering — end to end.
The founders brought the vision and the market insight — that Arabic-first AI tutoring was underserved in the Kingdom — and needed a partner who could take that from zero to a real product: the thesis, the feature set, the brand, the design language, and the engineering to bring it to life. Everything downstream had to be shaped from scratch, in close partnership with them.
The product challenge itself was compounding. A tutor genuinely bilingual from the ground up — not translated after the fact — needs to teach and speak naturally in both Arabic and English; it needs to be grounded in an actual curriculum so it stays in scope instead of free-associating; and it needs to autonomously drive its own teaching tools — pulling in questions, images, and assessments mid-lesson — without a human in the loop for every interaction.
Layered on top of that was a compliance requirement, not a nice-to-have: any platform teaching minors and processing their data in Saudi Arabia had to be built for PDPL, SDAIA AI Ethics, and NCA cybersecurity guidance from day one, with genuine visibility and control over what the AI was saying to every student.
We started at the very top of the stack, before a single line of code: shaping the product thesis, the feature set, the visual identity and brand, and the information architecture together with the founders — then built the engineering to match.
The platform runs on a bilingual, curriculum-grounded AI tutor operating in two modes: a session-based chat tutor and a real-time voice tutor over a full speech-to-text → LLM → text-to-speech pipeline. Both use tool calling rather than static scripting — the AI itself decides, mid-conversation, when to pull in a question, render an image, or launch a test, rather than following a fixed sequence. Before either mode reached a real student, we tested the Arabic experience directly with students in both chat and voice and iterated until feedback on both was strong.
Control sits with the humans running the platform, not just the AI: the Admin panel gives full, granular editorial control over what the tutor teaches from and how it responds — curriculum content, lesson structure, and every AI prompt are editable and versioned there, so the team can directly shape and correct the AI's output rather than treating it as a black box. On top of that, every AI response is logged with full visibility on our side, and every response is flaggable by parents through a simple built-in mechanism, so nothing the tutor says goes unreviewed. To meet PDPL requirements specifically, we built a dedicated PII anonymizer into the pipeline, so student data is stripped of identifying information before anything reaches the AI layer.
We also built with AI, not just for it. Coordinating Claude Code and Codex across the codebase let us move at a pace a traditional build process couldn't match — while a dedicated evaluation layer, with more than 500 automated evals running continuously against the tutor's own responses, kept quality honest as speed increased, catching regressions before they ever reached a student. That combination — AI-accelerated engineering with an AI-graded safety net — is what made a four-month, three-application, bilingual, voice-capable MVP possible in the first place. The stack itself ran on a fully containerized Docker setup deployed on our own VPS, with PostHog, Langfuse, and OpenRouter forming the observability and AI-infrastructure backbone for full-pipeline visibility from prompt to response to cost. And the whole engagement ran on a disciplined cadence — every decision, spec, and architecture note documented in Notion, with weekly meetings and sprint retros keeping founders and engineering in lockstep throughout.
In under four months, two first-time founders went from a strong market instinct to a working, bilingual, voice-capable AI tutoring MVP — product thesis, brand, design system, and full engineering stack built together from a blank page.
The AI tutor's autonomy — deciding for itself when to surface a question, an image, or a test — meant the product didn't need hand-authored scripts for every teaching path, while the 500+ eval layer meant that autonomy shipped with a safety net, not a leap of faith. The Admin panel gave the team real editorial control over the AI, not just oversight after the fact. Real students tested the Arabic tutor in both chat and voice, and the feedback validated the core bet: this could be bilingual by design and still feel natural in either language.
For two first-time founders, the outcome wasn't just an MVP — it was a fully shaped product, from thesis to brand to shipped platform, with the AI governance, compliance groundwork, and observability already built in for whatever came next. After delivery, the founders took the platform in-house to continue building on it.
Want similar results for your business?
In submitting this form, you are agreeing to Fynch’s Privacy Policy.
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Fast. Smart. Always on.
Fynch’s AI powered assistant is designed to help you unlock our full potential
Explore our product audit
Spot product and AI-readiness gaps early with our 80+ parameter audit, powered by McKinsey, Capital One, and RingCentral experience.
