PART 01 · OWN PRODUCT · LIVE 2025–

Deckio

AI that turns study material into a full exam-prep practice system.

FUNCTION: Source → review-ready practice + SRS
STATUS: Live, actively shipping
IN SERVICE: 2024 – present
ROLE: Solo · design → support

Deckio generating flashcards from an uploaded document — FIG. 01 — SOURCE → CARDS · GENERATION VIEW

Deckio spaced-repetition study session — FIG. 02 — SRS REVIEW · DUE-CARD QUEUE

01The problem

Students paste notes or upload a PDF and get nothing back they can actually study with. Generating cards is fast; building a practice habit with them is not.

Deckio converts arbitrary source material — text, PDF, DOCX, PPTX, image, YouTube — into review-ready content, then schedules it forward with spaced repetition so hard material resurfaces exactly when memory is fading.

02Approach

The generation layer runs Gemini Flash with per-mode prompt tuning — flashcards, long-form Q&A, and MCQ are separate calls with different density ranges. Large documents fan across multiple calls with deduplication rather than one giant context window.

Signed-in state lives in Supabase. A separate Python/Flask sidecar on Railway handles the binary work that can't run cleanly in serverless JS: Anki .apkg assembly via genanki, PyMuPDF page rendering, python-pptx extraction.

Deliberately no in-house ML inference, no vector store, no RAG pipeline. Gemini's context window is the retrieval layer — fewer moving parts, less to operate alone.

03How it works

User submits a source — text, PDF, DOCX, PPTX, image, or YouTube URL.
Flask sidecar extracts raw text; large PDFs are rendered to images for the vision path.
Next.js API routes the Gemini Flash call with mode, density, and language params.
Cards stream back; the user reviews inline or saves the deck to Supabase.
SM-2 schedules each card; the SRS queue resurfaces due cards daily.
Study Spaces wrap materials into an exam countdown with an AI topic path and a readiness model built from practice events.

04Capabilities

PDF / DOCX / PPTX / image / YouTube → cards
SM-2 spaced repetition, per-card ease tracking
Anki .apkg export — cloze + basic
Image-occlusion card editor
Study Spaces: countdown + topic path + heatmap
Math rendering — KaTeX inline & block
Mermaid diagrams in notes
Output-language picker with RTL rendering

05Stack

LANGUAGE: TypeScript (Next.js) · Python (Flask sidecar)
FRONTEND: Next.js 16 · React 19 · Tailwind CSS 4 · Framer Motion · Expo 55 / RN 0.83 (Android)
BACKEND / DATA: Supabase (Postgres + Auth) · Gemini 3 Flash · DB-backed rate limiting
SERVICES: DodoPayments · Resend · PostHog · Sentry
INFRA / DEPLOY: Vercel · Railway · EAS OTA · Supabase managed Postgres

06Engineering notes

Large-PDF vision path. Pages are rendered individually to images by PyMuPDF on the sidecar and sent as separate vision calls to stay under token limits — a 300s fluid-compute window, 10 images per call. A single 500-page request fans into ~50 sequential calls.
Quota consistency. Generation uses a DB-level consume-then-refund pattern (consumeUsageQuota + refundUsageQuota) so a timed-out Gemini call never silently drains a user's monthly allowance.
Profiles as billing source of truth. The subscriptions table is unused; all tier logic resolves from profiles.subscription_status + has_pro_limits + purchased_card_credits — collapsing pack buyers and trial-to-paid into a single row read.
SM-2 with Anki-style learning steps. New cards run a short-interval learning phase before graduating to the ease schedule; "Hard" ×0.8, "Easy" ×1.3, lapses reset to learning — matching Anki's real behaviour, not a simplified version.

07In service

1,500+REGISTERED USERS

40+PAYING USERS

2025LIVE SINCE

08What's next

Public deck gallery as an SEO / GEO acquisition engine for long-tail exam keywords.
iOS App Store launch — pending Apple developer enrollment.