Loading…
← HandbookLive
K53 Drill Master
An offline-first K53 learner's-licence drill PWA — 600+ DLTC-format questions in English, Afrikaans and isiXhosa, built for budget Android on 3G.
View the product pageThe problem
SA's K53 failure rate exceeds 60% — people fail on never having seen the question format under timed pressure, not on knowledge. Existing options (out-of-stock booklets, 15–40MB government PDFs, a 2009 quiz site) all fail the budget-device, prepaid-data user.
Architecture
How it’s wired
A routerless client SPA: all learning state lives in localStorage (including the SM-2 schedule), so the app works fully offline. Supabase only handles magic-link auth and the PayFast-ITN subscriber activation.
100
Lighthouse PWA score
~107KB
initial JS (gzipped)
R0.54
AI cost / month @ 10k calls
395
road-sign images
Stack
React 18 + ViteFramer MotionSupabase (auth + Edge Functions)OpenAI gpt-4o-minivite-plugin-pwa (Workbox)Capacitor (Android)PayFast
AI
The AI Tutor uses OpenAI gpt-4o-mini at temperature 0 with a localStorage cache — a deliberate, honest exception to the platform’s Anthropic default (it shows range, and keeps per-explanation cost near R0).
Data model
1 Supabase table + 5 localStorage stores
- subscribers (RLS: read-own, insert service-role only; renewals via upsert)
- k53_sr_v1 — the SM-2 spaced-repetition card state
- k53_usage / k53_premium / k53_history_v1 / k53_streak_v2
Key decisions
Why it’s built this way
ADR-01
Routerless SPA
- Context
- 11 game modes + 5 tabs looks like it needs a router.
- Decision
- One activeGame state machine, props + localStorage, no Redux/Zustand.
- Consequence
- −50KB; the Android back button was then fixed with a 10-line pushState/popstate trick.
ADR-02
PayFast over Stripe
- Context
- Stripe needs Atlas for ZAR cards; the user pays with Capitec/FNB.
- Decision
- PayFast hosted checkout; the passphrase doubles as the HMAC key for the ?unlock token.
- Consequence
- Native ZAR, zero extra secret management — but SA-only.
ADR-03
gpt-4o-mini, temperature 0, cached
- Context
- Per-question explanations needed, but GPT-4 at scale exceeds revenue.
- Decision
- Mini at temp 0 (identical question → identical answer → cache hit), djb2-hashed localStorage cache.
- Consequence
- ~R0.54/month at 10k uses.
The hard parts, solved
Spaced repetition from scratch (SM-2)
Each question is a card {interval, easeFactor, repetitions, nextReview}. Correct answers walk an interval ladder (1d → 6d → round(interval × EF)) and adjust ease; wrong answers reset. getDueIds() returns the most-overdue first; "mastered" = interval ≥ 21 days. Powers the Weak Spots tab with zero manual config.
Road-sign extraction
395 sign images pulled from the official manual PDF via paint-operation counting to distinguish real signs from page chrome, with a runtime SVG fallback so a corrupted image never shows a broken state.
Load-shedding-proof audio
Sounds are synthesized at runtime via the Web Audio API (zero audio files, offline-safe), with ctx.resume() on every call to satisfy the iOS gesture requirement.
Sourced from nanda-k53-drill-master — BUILDLOG_k53, WONDERLAND_DECISIONS_K53, spacedRepetition.js.
.jpg&w=3840&q=75)