AuditFlow - Observability Layer for AI Review Pipelines

Turn Dify-style AI review from a black box into a debuggable, replayable and swappable pipeline for Java backend teams.

Problem

Teams embed Dify into Java B2B systems for merchant onboarding and decision-document review, but complex flows explode into unmaintainable Dify nodes. Dify lacks a CLI/MCP, so AI can't debug it directly, and every business system has to re-implement auth, retries and fault tolerance. Worse, many long-tuned workflows work worse than just sending documents to a local Claude Code and letting it reason freely.

Solution

A middle-layer review orchestrator: break review rules into versioned, unit-testable steps, expose a CLI and MCP endpoint so coding agents can debug directly; log every input, model call, intermediate conclusion and final decision for replay and A/B model swaps; business systems call one stable review API while the middle layer handles auth, retries and fault tolerance.

Users

Small and mid-size enterprise dev and ops teams on Java stacks that are bringing AI into review workflows.

MVP

Review-step DSL with versioning; CLI for single-step debugging and an MCP server; audit log and replay UI; pluggable model adapter; REST API for existing Java systems.

Revenue

Open-core with a paid enterprise edition priced per review volume or seat; one-time private deployment license; hosted version billed per API call.

Why now

Many traditional B2B teams are moving from Dify prototypes to production but lack an observability and debugging layer; AI agents can now debug workflows themselves, creating a window for indie developers to build a lightweight middle tier.