CapyData

CapyData turns everyday learning into a World ID-verified game: you raise a virtual capybara by feeding it snippets of your life—documents, images, trivia answers—while the same interactions label data that improves AI. The app pairs a modern Next.js 15 + FastAPI stack with Worldcoin’s Mini App rails to guarantee every label comes from a real human and every reward goes back to that human.

Description

CapyData is a “learn-to-earn” virtual-pet platform that combines three proven ideas:

1. Tamagotchi-style engagement keeps users returning to care for a digital companion.
2. NotebookLM-like personal knowledge bases let users upload or author content and query it with AI.
3. Scale-AI-grade data-labeling workflows, but crowdsourced through short, fun games rather than paid task queues.

User Journey

• Onboard with World ID
  One-person-one-account login prevents Sybil attacks and keeps labels honest while hiding the user’s real identity with zero-knowledge proofs.

• Adopt a capybara
  The pet’s health and appearance reflect the user’s activity—XP from games, new insights generated, and data contributions.

• Play micro-games
  Four launch titles—Image-Quality Judge, Language Flashcards, Mood Detective, Knowledge Trivia—convert image rating, translation checks, sentiment tagging, and fact verification into enjoyable 30-120 second rounds. Academic studies show gamification sustains data-labeling quality at scale.

• Feed data
  Users can add data sources from the Internet and from personal social accounts to enable personalized insights in the future.

• Receive insights
  A GPT-4o-powered assistant digs into the user’s personal corpus and game history to surface strengths, blind spots, and tailored learning goals.

• Explore AI studio
  Provides detailed analytics on submitted data and chat interactions, inspired by best-in-class AI web2 features.

• Earn & spend
  Points unlock cosmetic skins and, in the next milestone, on-chain NFTs and marketplace revenue share.

Why World ID Matters

Traditional crowdsourcing fights constant bot infiltration; requiring an orb-verified World ID ensures every label originates from a living human, improving data reliability and enabling fair reward distribution without leaking personal data.

Long-Term Vision

The same personal knowledge graphs that answer “What did I read last week about zero-knowledge proofs?” will become licensable datasets in an open marketplace. World ID and RLS (Row-Level Security) rules in Supabase guarantee both attribution and rev-share as that data trains future models.

CapyData — Tech Stack Overview

Front-end

• Next.js 15 gives App Router, React Server Components and Turbopack for 20 % faster dev builds.
• React 19 RC provides streaming UI and built-in async error boundaries, which we use for game loading states.
• Tailwind CSS 4 supplies a lightweight pixel-art theme; Radix Primitives standardise modals and tooltips.
• A PWA manifest lets users add CapyData to the World App home screen with offline caching for static assets.

Back-end

• FastAPI 0.111 exposes JSON endpoints; the new fastapi dev CLI sped up local iteration during the 48-hour build window.
• Python 3.11 workers call the OpenAI API for chat, function calls and embeddings; GPT-4o is the default model.
• Notte SDK supplies lightweight embedding pipes when we need to stay entirely on-prem for private corpora.

Data Layer

• Supabase Postgres with Row-Level Security ensures each user reads and writes only their own pet, game and knowledge rows—even if a token leaks.
• A single game_config.ts file in the repo holds metadata (route, rewards, colour theme) for every mini-game, hot-reloaded by both the hub page and the API to guarantee consistency.

Partner Tech & Benefits

Worldcoin MiniKit: Web-view bridge & World ID auth | Instant access to millions of verified humans, Sybil resistance

Hacky but Notable

• Edge-deployed embeddings. Because World App web-views restrict heavy downloads, we compress OpenAI embeddings with PCA to 128 dims on the fly, then store them in a Postgres vector column (extension pgvector)—shrinking storage 4 × while keeping recall above 0.92 on our test set.
• Game-driven AI eval loop. Every trivia answer is dual-logged: once for XP, once in an evaluation table that automatically scores GPT-4o’s own answer to the same question, providing continual supervised feedback with no extra user effort.
• Zero-config split builds. We piggy-back Next.js Turbopack’s incremental compiler to produce both a normal web bundle and a reduced “mini-bundle” (<200 kB) that World App downloads after orb-verified login, shaving first-interactive to 1.1 s on mid-range Android.

Together these pieces create a fast, privacy-preserving playground where people, pets and models all learn from each other—and every contribution is both verifiable and rewarded.