Oynix Guardian

AI writes the code. Guardian decides if it's safe to spread.Enterprise AI Governance for Multi-Agent Development

Comment

Inspiration

AI agents now write code, docs, tests, and answers. But every pull request silently raises questions nobody answers: which services break, which docs go stale, which AI agents are now operating on outdated context, and is it safe to let this change spread? Today these decisions are manual or ignored.

What it does

The moment a pull request is opened against main, UiPath Maestro runs a case:

  1. Analyze Impact — Oynix Guardian fetches the PR's changed files from GitHub, walks the dependency graph, and computes a risk score from the affected services, docs, and active AI agent sessions.
  2. Decision — low-risk changes auto-propagate; high-risk changes pause and raise a UiPath Action Center approval ("Risk 82/100 — approve?").
  3. Propagate — once a human approves, Guardian updates the graph, regenerates the affected docs, and notifies the AI agents. The case closes.

It gates the change before it merges — preventing bad propagation, not reacting to it.

How we built it

  • guardian-service (Node/Express + Neo4j) parses the real TypeScript imports, doc references, and each AI agent's declared depends_on, builds a dependency graph, and computes a deterministic, explainable risk score. A /impact-pr endpoint fetches a PR's changed files from GitHub directly.
  • UiPath Maestro Case — 3 stages (Analyze Impact → Human Approval → Propagate) with a human-in-the-loop gate and exception handling.
  • UiPath API Workflows call Guardian over the unified HTTP connector; UiPath Action Center hosts the approval.
  • Runs on UiPath Automation Cloud. Built with Claude Code via UiPath for Coding Agents.

Example

A PR changing shared-sdk/auth.ts → 4 downstream services, 3 active AI agents on stale context, 4 docs to regenerate → Risk 82/100 → HIGH → human approval required. A PR changing a leaf service → 10/100 → LOW → auto-propagates. The numbers come from the graph, not a constant.

Challenges

Designing a risk model that's real and explainable (every point traces to a graph edge), and making the orchestration a genuine case — branching, human handoff, exception escalation — rather than a linear script.

Accomplishments

A working, PR-triggered governance pipeline where UiPath is the orchestration and governance layer, with a live human-in-the-loop gate driven by real impact data computed from the actual PR.

What's next

Wire Guardian to an org's real knowledge sources (the full Oynix engine), expand the risk model with historical incident data, and support multi-repo blast radius.

Built With

Share this project:

Updates