Lethe

Lethe is a privacy-first medical billing audit system that helps patients detect and dispute costly errors without exposing sensitive data. Users upload a hospital bill, and Lethe immediately strips all personally identifiable information before any AI processing occurs. Three independent AI agents analyze the redacted bill in parallel, then share findings and re-evaluate them through a second round of peer review. Only results with 2 of 3 agreement are accepted, ensuring high-confidence error detection.

Medical billing errors cost Americans over $100B annually, yet most patients lack the time or expertise to challenge them. Lethe directly addresses this by identifying overcharges, incorrect codes, and duplicate fees, then automatically generating a structured appeal letter for the user to review and send.

The original bill is never stored, and all sensitive data is wiped from memory after processing. For verification, Lethe anchors a hash of the audit and its outcome on-chain, creating a tamper-proof record without revealing private information.

By combining multi-agent AI consensus, zero data retention, and verifiable blockchain records, Lethe makes it simple to recover money lost to billing errors while preserving complete user privacy.

Lethe is built as a full-stack system combining a React/Next.js frontend with a FastAPI-based orchestration backend that coordinates parsing, redaction, AI agents, and on-chain actions. When a user uploads a bill, a deterministic parser extracts structured data, followed by a multi-layer redaction pipeline (regex + LLM sweep) that removes all PHI before any model interaction.

The core of Lethe is a 3-agent consensus system using independent LLM providers (OpenAI, Anthropic, and Google). Each agent runs in parallel, then communicates findings over a peer-to-peer mesh (Gensyn AXL) to enable a second round of reflection before reaching a 2-of-3 quorum. This avoids single-model bias and creates a more reliable audit layer.

For persistence and verification, Lethe uses the 0G stack: 0G Chain to anchor audit hashes and outcomes, 0G Storage for encrypted audit blobs and rulebooks, and optionally 0G Compute for decentralized inference. A Node.js sidecar handles signing and interaction with 0G services, enabling trust-minimized compute and storage without exposing raw data.

KeeperHub is used to trigger real-world actions via on-chain workflows, including anchoring audits, recording disputes, and attesting when appeals are sent. These workflows interact with a unified smart contract deployed on Ethereum Sepolia.

Some notable “hacky” elements include running three independent AI agents with separate Dockerized P2P nodes for real message passing, a fallback system that dynamically switches between decentralized and centralized inference, and a zero-retention memory model that wipes all bill data immediately after processing while still preserving verifiable proofs on-chain.