← IndexWork / 2024 — 2025

stella AI

Private-cloud model deployments with hardened backend infrastructure, built for teams whose internal strategy cannot pass through a public provider — a position that aged well as courts began compelling providers to retain deleted conversations.

Paused

Role: Founder
Year: 2024 — 2025
Stack: Next.js, TypeScript, Private Cloud, LLM Infrastructure

The problem

Companies wanted to put AI into their workflows. They also could not afford for their internal strategy, their contracts, or their customer data to sit in a third party's logs.

That was treated as paranoia for a while. Then a US federal court ordered OpenAI to preserve and segregate all ChatGPT output data — including conversations users had deleted, and conversations held in "temporary" mode. Deleting a conversation had never guaranteed it was gone.

The concern was not paranoia. It was just early.

The approach

stella AI deployed models into private cloud, with the isolation guarantees written into the infrastructure rather than into a terms-of-service page:

Private VPC only, egress denied by default
No prompt or completion retention
Customer-managed keys

Tenancy was not a feature of the product. Tenancy was the product; everything else was packaging.

Status

Paused. The thesis holds — arguably better now than when we started — but Nora had more pull, and running both properly was not possible.

From the source

infra/isolation.ts

/** Tenancy is the product.
 *  Everything else is packaging. */
export function provision(tenant: Tenant) {
  return createInferenceCluster({
    tenantId: tenant.id,
    network: "private-vpc-only",
    egress: "deny-all",
    retention: { prompts: "none" },
    keys: tenant.customerManagedKeys,
  });
}