Megh Vyas
Backend · Infra · SRE
← All work

AI governance

REX

AI pipelines you can draw, run, and be held accountable for, with guardrails and audit trails built into every node.

The accountability gap

Teams ship AI workflows they cannot explain or govern. When a pipeline makes a bad call there is no record of what ran, who was allowed to run it, or what guarded the input and the output. 'Responsible AI' stays a slide, never a system.

The graph

01

Built visual DAG authoring over a Fastify control plane, executing pipelines as directed acyclic graphs on a BullMQ worker runtime.

02

Added guardrail nodes (input-guard, output-guard, json-simplify), so every pipeline validates what goes in and what comes out.

03

Put it all under RBAC and attribute-based authorization with workspace sharing, per-workflow permissions, and IAM policies.

04

Made behavior configurable at runtime with a domain-config resolver layering global, user, and workflow overrides, plus a model registry and hyperparameter experiments.

05

Wired governance in as a first-class concern: KPI and observability endpoints, alerting rules, and GDPR operations for export, deletion, consent, and retention.

Governed by default

A platform where AI workflows are visual, permissioned, measured, and compliant by construction. When something goes wrong, the answer to what happened and who could do it is already in the system.

0

apps: api, worker, ui

DAG

authored visually

GDPR

built into the core

Wired in

Visual DAG authoring + BullMQ execution runtime
Guardrail nodes: input-guard, output-guard, json-simplify
RBAC + ABAC with workspace sharing and IAM policies
Runtime domain-config resolver (global / user / workflow)
Model registry + hyperparameter experiments
GDPR: export, delete, consent, retention, audit