Reading Platform

The palette is declared once in globals.css as CSS custom properties and mapped into Tailwind 4 through an @theme inline block.Tailwind utilities, shadcn primitives and Radix components share the same values: #5AA6F2 for primary, #FFD84D for accent, #7EDCA0for success and #FF6B6B for alerts, over a sky gradient from #F9FAFB to #E8F4FF behind every screen. A complete dark set is definedalongside it.

Radius is one 8px base, and every step is derived from it with calc, from 4px on the smallest control to 24px on the large cards; rounded-full, 3xl and 2xl are what the interface reaches for most. Fredoka at 400, 500 and 700 is self-hosted through @fontsource and used forheadings and body alike, which is what gives the product its soft, child-facing voice.

Reedo gives the interface its tone through expressive mascot states, drawn as vector art and built as React components with sharedproperties. The character adapts to dashboard headers and sign-in screens, with an optional hover effect that gently lifts and enlarges it.

The watching state follows the pointer or a finger on a tablet, with movement constrained so the gaze stays within the eyes. Each statehas its place: watching on sign-in, happy in the header, stars during reading and celebrating on the session summary, with additionalstates held in reserve.

The reading session captures the microphone at 16 kHz mono and sends Int16 frames over a WebSocket to the speech-scoring service,in a continuous stream. Interim transcripts arrive while the child is reading; final ones arrive per utterance, and a reading score followsonce the service has run forced alignment on the passage.

Final transcripts are aligned to the passage by character-level edit distance across the whole accumulated transcript, which is what keepsthe highlight in place when a child repeats or skips a word. The same audio is encoded to FLAC in the browser and uploaded to S3 througha pre-signed URL the API issues with the session.

Each reading starts with the API minting an RS256 session token whose issuer, audience and expiry the speech service verifies, and whosesession identifier encodes the assignment, pupil, exercise and attempt. The web client opens the speech socket with that token, so theservice can attribute every score without trusting the browser.

Review selection is weighted against performance. calculateReviewWeight subtracts mean accuracy from 100 and adds a ten-point bonus while a child has two or fewer attempts on record, so an exercise averaging 60 percent carries a weight near 50 while one averaging 95 percent carries 15. Selection then walks the weighted total.

The result comes back as an HS256 token the API verifies before anything changes. Accepting it then updates stars with an atomic delta, rolls up study-plan progress, extends the reading streak, clears the class report cache and evaluates achievements, in that order, and theend-to-end suite exercises that path with signed results against a real Postgres.

The API uses NestJS 11 and Postgres 17 through TypeORM, with routes versioned under /api/v1 and documented through Swagger.Role-based access covers administrators, school administrators, teachers, parents and pupils, checked by a RolesGuard. Every servicescopes its queries with shared helpers that return a 404 for records outside the caller’s scope.

Teachers build ordered study plans from the passage library and assign them to a class or to one pupil, and parents get the same toolsthrough a personal group. Class reports are cached as jsonb with a configurable time to live and invalidated whenever membership orprogress changes, while role-sensitive fields are shaped by class-transformer groups so one endpoint serves a teacher, a parent and apupil different views of the same record.