Relayer
AI runs it. You don’t.
About
The problem it solves **Today's AI agents can think, but they can't _do_.** They can research, summarize, and plan — but the moment a task requires someone to physically show up, build something, or interact with the real world, they hit a wall. Meanwhile, hiring humans for complex jobs is a fragmented, trust-dependent process riddled with coordination overhead, unclear scope, and payment disputes. **Relayer bridges the gap between AI intelligence and human capability** by creating a self-sustaining AI agent that autonomously coordinates real-world work. ### For Clients (Job Posters) - **Turn vague ideas into structured plans.** Instead of figuring out every step yourself, describe what you need in plain language. Relayer's AI clarifies your intent, then decomposes the job into sequential, actionable subtasks — mixing AI research with human physical work. - **Trustless payments with on-chain escrow.** Your ETH is held in a smart contract and only released when work is verified. No middlemen, no disputes, no chargebacks. - **AI-powered quality assurance.** Every proof submission is analyzed through multi-pass AI vision verification — checking for fraud, relevance, completeness, and cross-task consistency — so you don't have to manually review every deliverable. ### For Workers - **Get matched to work you're good at.** Skill-tagged tasks and a recommendation engine surface the jobs that fit your expertise — no endless scrolling through irrelevant listings. - **Guaranteed payment on approval.** Funds are locked in escrow before you start. Complete the work, submit proof, and get paid automatically on-chain. - **Build portable reputation.** Your verification scores and completion history build a reputation tier (Bronze → Silver → Gold) that earns you priority access and bonus payments up to +10%. ### For the Ecosystem - **A new primitive: the self-sustaining agent.** Relayer maintains a 20–30% profit margin on every job to fund its own compute, gas, and API costs. It tracks its own P&L, sells AI services via x402 for additional revenue, and operates with full financial transparency on a public dashboard. This isn't an agent that _costs_ money to run — it's an agent that _earns_ money to survive. - **Hybrid AI-Human workflows, on-chain.** AI handles what it's best at (research, analysis, verification), humans handle what they're best at (physical tasks, creative work, local presence), and smart contracts handle what neither should be trusted with (money). - **Verifiable, transparent coordination.** Every action the agent takes — decomposing jobs, assigning tasks, verifying proofs, paying workers, withdrawing profit — is logged on-chain or streamed in real-time via the agent dashboard. Nothing happens in a black box. Use of AI tools and agents Relayer's core is an autonomous AI agent built on GPT-4o that manages the entire lifecycle of a job — from understanding what the client wants, to planning the work, executing research, and verifying deliverables. **Clarifier** Converses with the client to resolve ambiguities before job creation, asking focused follow-up questions and showing a live task preview each round. **Decomposer** Breaks the finalized job into a sequence of AI and human subtasks, tags each with skills for worker matching, and enforces a 20–30% profit margin. **AI Executor** Runs research tasks before human work begins, producing key facts (pricing, regulations, specs) that get injected into human task descriptions. **Verifier** Multi-pass AI vision pipeline that checks submitted proof for fraud, requirement compliance via semantic reasoning, and cross-task consistency with previous deliverables. **Orchestrator** Listens for on-chain events, triggers the appropriate AI pipeline, executes blockchain transactions, tracks costs/revenue, and streams all actions to a public dashboard. **x402 AI Services** Sells standalone AI capabilities (image analysis, text classification, photo verification) through paid API endpoints on Base, generating revenue independent of job profits.
How it's made
Challenges we ran into ### Making AI Verification Fair (Not Just Strict) Our earliest verifier was a single GPT-4o call with a simple prompt: "does this image match the requirements?" It rejected almost everything. A worker who was asked to photograph a "dog at a park" submitted a golden retriever — and the verifier failed it because "golden retriever" didn't literally match "dog." We split verification into three separate passes (fraud detection, requirements check, cross-task consistency), each with its own prompt emphasizing semantic reasoning over literal string matching. Getting the weighted confidence formula and per-score floor thresholds right took dozens of iterations before it stopped being either too lenient or too harsh. ### Sequential Task Dependencies & Orchestration The smart contract enforces that Task 2 only unlocks after Task 1 is approved, but the agent posts all tasks in a single batch at job creation time. Coordinating sequence indices between the decomposer, the chain, and the orchestrator's local state was a constant source of off-by-one bugs — mismatches meant jobs would complete early or never finish. On top of that, AI research tasks run before anything hits the chain, and their findings need to be injected into the right human task descriptions. We built a two-phase pipeline: Phase 1 executes AI tasks and collects key facts, Phase 2 reads the next task ID from the contract and posts human tasks with relevant AI context stitched in. Getting the handoff between phases reliable — especially when an AI task fails or the RPC call times out — took significant iteration.