ARCA

Demo

About

ARCA is an AI-powered outsourcing platform that turns company projects into structured tasks, records how people complete them with AI, and distributes rewards based on qualified contributions. Companies can choose the results that best fit their needs, while workers can earn rewards for meeting clearly defined requirements. When a company submits its goals and requirements, AI breaks the project into tasks based on the existing project context. This includes defining function interfaces, input/output JSON schemas, file requirements, and acceptance tests. These shared contracts help independently produced deliverables fit into the existing project. Once the company approves the plan, tasks are published, and participants work against the same requirements and completion criteria. Workers use a transparent desktop panel alongside their existing development tools to perform tasks with the platform’s AI. ARCA records the agents and models used, prompts, tool calls, file changes, and AI usage costs. This addresses a growing challenge in AI-assisted outsourcing: receiving code alone does not explain how or why it was created. Connecting deliverables to their work history supports review, accountability, and handover. ARCA uses World’s proof of humanity to support fair participation. It restricts the same person from entering the same task through multiple wallets and connects protected submission actions to human approval. Companies review submissions against predefined rules and qualification criteria, then select the result that best meets their needs. Rewards follow a differentiated distribution model rather than a winner-takes-all approach. Qualified contributors can participate in the reward pool even when their submission is not the final selection, with allocations varying by ranking. AI usage costs are accounted for within each worker’s gross reward allocation. Curvegrid connects payout plans with supporting evidence records, while Sui handles escrow-based USDC settlement. Together, these components connect project requirements, AI-assisted work, human review, and reward settlement into a traceable outsourcing workflow.

How it's made

We built ARCA as four connected components: a web platform for companies and workers, a desktop panel for AI-assisted work, a backend for task management and settlement, and blockchain adapters for partner integrations. The application uses TypeScript, a Next.js web interface, and Prisma to manage projects, tasks, participants, submissions, AI execution history, and settlement records. We separated the AI workflow into project analysis, task decomposition, execution, and review. During decomposition, the system produces structured requirements alongside natural-language instructions, including function interfaces, input/output schemas, permitted files, and acceptance criteria. Execution and evaluation reference the same task contract, making it possible to compare submitted work against the original requirements. The desktop panel lets workers interact with platform AI while using their existing development environment. Each execution records its model and agent, prompts, tool calls, file-change hashes, usage, and costs. These records are associated with the relevant task and submission, while proposed changes pass through human review and approval. One distinctive implementation detail is that client-submitted work history is checked against server-side AI execution records instead of being trusted on its own. For World, we integrated an IDKit-based proof request flow with server-side verification. Verified human identity is associated with task participation to restrict duplicate entries by the same person using different wallets. We deliberately separated wallet ownership from proof of humanity: wallets support accounts and transactions, while World supports person-level eligibility and protected submission approval. On Sui, we implemented Move smart contracts for escrow and settlement. Company deposits are associated with task budgets, and finalized allocation plans determine worker rewards, AI cost allocations, and refunds. The contracts include checks against duplicate settlement and spending beyond the available task budget. This makes reward execution verifiable through onchain transactions rather than only through changes in an application database. We integrated Curvegrid MultiBaas with a Solidity settlement-record contract. The backend records payout-plan and evidence hashes, then links the resulting Sui settlement receipt. MultiBaas provides the contract-call and transaction-query interface for this EVM audit layer. Prompts and private work files remain offchain; onchain records contain hashes and settlement information needed for verification. A significant engineering challenge was maintaining consistency between work evidence, reward calculations, and transactions across two chains. The EVM record and Sui payment are not a single atomic transaction. We therefore persist execution states and transaction receipts so that failed or uncertain operations can be reconciled before retrying. This adds coordination overhead, but helps prevent duplicate execution and makes settlement failures traceable. We validated the implementation through automated tests and separate partner integration checks. These covered application workflows, authorization, reward calculations, World verification through the official simulator, Curvegrid testnet writes and readbacks, and Sui testnet escrow transactions. We treated these as separate validation milestones rather than presenting individual integration checks as proof of a completed end-to-end USDC competition.