AI Hub

One governed path to enterprise LLMs

Every model interaction takes one governed path: access brokered centrally, inputs checked, outputs logged, so operators can see and control how models are used.

Platform engineer & builderAI governancePythonAWS

Harrison Halperin

Project brief

Shipped

Problem

Ad hoc model access scatters credentials, data handling, and audit decisions across engineering teams.

Outcome

Every model interaction takes one governed path, access brokered centrally, inputs masked, and a ledger entry for each request.

Evidence

The load-bearing piece is the request pipeline: every request is validated, masked, evaluated, and recorded.
JWKS-validated tokens and a centralized PII vault sit in front of every model call.

Key decisions

Frontend zone: The Agent Admin Center never calls model providers directly, registry and audit go through the APIs.
Gateway zone: JWKS-validated tokens and a centralized PII vault sit in front of every model call.
Ledger zone: Every request, response, and decision writes to a durable audit ledger keyed by trace ID.
Infra zone: Terraform Cloud overlays provision each environment from development through production.

Field record

The load-bearing piece is the request pipeline: every request is validated, masked, evaluated, and recorded. I built the platform as a coordinated set of services and Terraform-managed infrastructure rather than one app: an Agent Admin Center for registry and approvals, a registry API with a server-side identity-token proxy, and a core API that owns the PII vault, LLM gateway, and audit ledger, so new processors and model providers slot in cleanly. MeTime explores a related many-surfaces idea on the personal side. Spawn engines and one service layer, many surfaces are the sibling habits. The goal is one governed front door instead of ad hoc API keys in scripts.

Routing manifest

The first figure follows one request through the governed path from entry to ledger. The second shows how the admin UI, APIs, and infrastructure divide responsibility under one governance spine.

The governed request pipeline

Every model call is validated, masked, invoked through a gateway, and recorded before a response returns, so no interaction bypasses the ledger.

Request

client or gateway

enters the core API

AuthGate

JWKS · identity proxy

validates the bearer token

PII + safety

mask · dehydrate · classify

sanitizes the prompt

Model invoke

AI gateway · provider

processes the request

LedgerGate

record + audit

every interaction persists

Response

returns to caller

Governance spine

One governance spine, several subsystems

The admin frontend and backend APIs are split by concern. All of them read and write through the same governed data layer, provisioned by one Terraform substrate.

Hub

Governance spineauth · PII vault · ledger

shared by every API

Agent registryadmin UI + registry API

onboarding, approvals

LLM gatewaycore API

model invocation

Audit ledgercore API read path

review by trace ID

Terraform substrateEKS · IAM · DynamoDB

provisions every environment

Zone notes

Frontend zone

The Agent Admin Center never calls model providers directly. Registry and audit go through the APIs.

Gateway zone

JWKS-validated tokens and a centralized PII vault sit in front of every model call.

Ledger zone

Every request, response, and decision writes to a durable audit ledger keyed by trace ID.

Infra zone

Terraform Cloud overlays provision each environment from development through production.

Skill manifest

AI governance

Model calls go through one brokered path with input checks and logged outputs.

Python

Hub brokerage and policy checks are Python services on the shared model path.

Terraform

Hub infra changes ship as reviewable Terraform beside the brokerage services.

Platform engineering

One brokered model path is the paved default instead of per-app sprawl.

Security

Input checks and access brokerage assume both agents and humans will make mistakes.

AWS

AI Hub runs on AWS primitives: compute, data, and IAM treated as coded infrastructure.

Cloud automation

AI Hub codifies cloud operations so provisioning and remediation stay repeatable.

Agentic

Agents call models through the hub; operators and policy decide what is allowed.

CI/CD

Hub changes go through checks and merge gates like any other platform service.

Angular

Angular structures the application UI for AI Hub.

Kubernetes

AI Hub deploys workloads on Kubernetes with an explicit operational surface.

DynamoDB

AI Hub stores keyed data in DynamoDB matched to real access patterns.

Amazon Bedrock

AI Hub runs foundation models through Bedrock with explicit model and eval hooks.

FastAPI

Governed LLM gateway runs as a FastAPI pipeline with PII sanitization and rehydration.

Portfolio graph

Hover nodes · click empty canvas to zoom/pan

Maps how the work here connects. Open the Portfolio graph →

Professional
Personal
Skill
Writing