XRPmodels

Decentralized AI inference marketplace on XRPL — miners run LLMs inside TEE VMs, clients pay per token via on-chain escrow, and ZK proofs guarantee no model substitution or hardware fraud.

Demo Video

Project Information

At a Glance

Decentralized AI inference marketplace on XRPL — miners run LLMs inside TEE VMs, clients pay per token via on-chain escrow, and ZK proofs guarantee no model substitution or hardware fraud.

Description

The Problem

AI inference today is a black box controlled by centralized providers. You send a prompt, you get a response, and you trust that the right model ran, that your data wasn't logged, and that the price was fair. There is no way to verify any of it. There is no open market. There is no competition on price or quality. And there is no role for a decentralized ledger — because until now, there was no way to cryptographically prove what happened inside a GPU.

XRPmodels changes that.

What XRPmodels Is

XRPmodels is a decentralized marketplace where miners sell LLM inference and clients buy it — with every payment settled natively on the XRP Ledger and every computation backed by a zero-knowledge proof.

  • A miner downloads a VM image, launches it on TEE-capable hardware, and starts earning OMC. The VM handles everything: key generation, on-chain registration, inference serving, and payment settlement.
  • A client picks a model from the live on-chain registry, locks funds in an XRPL escrow, and starts prompting. When the session ends, the escrow settles on-chain — the client pays only for tokens consumed, unused funds refunded automatically.

No platform fee. No API key. No account. Just a wallet.

Why This Matters

XRPL becomes the settlement layer for AI compute. Every LLM session generates real, metered XRPL transactions — escrow creation, escrow finish, refund payments, miner payouts. This is not a token launch hoping for speculation. It is a utility that produces on-chain volume proportional to real-world AI usage. As AI demand grows, XRPL transaction volume grows with it.

First real use of XRPL escrows for metered service payments. XRPL escrows have existed since 2017 but are almost exclusively used for time-locked holds. XRPmodels repurposes them as a trustless prepayment mechanism — the escrow settled with a ZK proof attached as a memo. A new pattern any XRPL project could adopt for metered services: bandwidth, storage, compute, API calls.

ZK proof verification via Groth5 precompiles. Every escrow settlement includes a proof that the VM ran the exact model the client requested, inside genuine TEE hardware, with encrypted GPU memory. This is the first application using Groth5 precompiles to enforce computational integrity directly on the XRP Ledger.

Permissionless supply side. Anyone with TEE-capable hardware can become a miner. No approval process, no staking requirement, no geographic restriction. The ZK proof replaces trust — if the proof verifies, the miner is legitimate.

How It Works — Zero Ambiguity

  1. Miner boots VM → wallet generated inside TEE → ZK proof (Groth16 via RISC Zero) binds model hash, TEE platform, wallet address → NFT minted on XRPL registry.
  2. Client queries registry → picks model → handshake with VM → receives wallet, price per token, ZK proof → verifies.
  3. Client creates EscrowCreate locking OMC to VM wallet with DestinationTag binding to the session.
  4. VM detects escrow on-chain → serves inference via OpenAI-compatible API → meters token usage.
  5. Session ends → VM submits EscrowFinish with ZK proof in memos → Groth5 precompile verifies → escrow releases → refund to client + payout to miner. All on-chain. All auditable.

https://ismael-f00607c9.mintlify.app/ https://docs.google.com/presentation/d/1LyrVTtLyf9eR0YjQhiQUPbPytXJOvVk7haUNRuzhCr0/edit?usp=sharing

Technical Details

Two-container Docker Compose per miner: llama-server (llama.cpp, OpenAI-compatible) + miner-vm (Python FastAPI). No central backend — clients connect via Go CLI or React UI directly to XRPL nodes and miner VMs.

Attestation: TEE report → RISC Zero guest verifies cert chain → Boundless proving market (Base) returns Groth16 seal + journal binding VM image hash, TEE platform, GPU status, wallet address.

Settlement on groth5.devnet: Client locks OMC via EscrowCreate with DestinationTag. VM polls account_objects, serves inference, settles via EscrowFinish with ZK proof memos verified by Groth5 precompiles. Refund + miner payout as real Payment txns.

Team

3
AS

Andrii Syvash

IM

Ismaël MEHDID

JM

Jan Matter

Hackathon

HACK THE BLOCK 2026 Paris Blockchain Week XRPL Hackathon

Duration

Apr 11, 6:30 AM - Apr 12, 6:00 PM UTC

View Hackathon Details