Tasmi: AI-powered Quran practice app

Tasmi helps students practice Quran recitation at home with honest word-level feedback on the printed mushaf page. Anyone can try without an account. Parents and teachers can see whether the week's assigned lesson was actually practiced.

Story

I built Tasmi because unsupervised Quran practice at home is easy to fake and easy to practice wrong. Parents cannot sit through every repetition. Teachers cannot hear whether homework happened. Most Quran apps optimize streaks for the individual. Tasmi is AI-backed homework between live lessons.

The practice surface is the mushaf page itself, live at tasmi.app. After a try, matched and missed words highlight in place. Tajweed colors are a reading aid, not a score. Anyone can try a passage without signing up.

Generic speech APIs were a poor fit. Recitation is a narrow domain, audio should not leave my infrastructure, and hosted models drifted on this material. I collaborated with the person who trained a recitation-specific speech model, then designed and assembled dedicated hardware to run it in production.

The product is in production for anonymous practice and the household homework workflow. Paid billing is still rolling out.

Word-level AI feedback on the Madani mushaf page, highlighting matched and missed words in place.

Dedicated AI inference server I designed and assembled to run Tasmi's recitation-trained speech model.

Impacts

Launched public AI practice at tasmi.app so anyone can recite a passage and see what to fix without creating an account.
Partnered with the collaborator who trained a recitation-specific speech model and shipped it as Tasmi's production inference path.
Designed and assembled dedicated inference hardware so the model runs privately, without a third-party speech API.
Practice and highlights sit on the printed mushaf page so missed words appear where a student would look on paper.
Parent and teacher workflows show whether the week's assigned lesson was practiced at home.

Challenges & Solutions

Challenges

Generic hosted speech APIs were the wrong cost, privacy, and accuracy shape for Quran recitation.
Students memorize from a printed page, so a scrolling text app would make the feedback hard to trust.
Live recitation is long, uneven, and mixed with household noise.
Parents, teachers, and students need different views of the same practice.

Solutions

Brought a collaborator's recitation-trained model into production on inference hardware I built, under my control.
Show practice and highlights on the mushaf page they already use.
Built a capture-to-feedback path that turns a real try into word-level results students can act on.
Separate household and teacher workflows around one shared practice mechanic.