Companion AI

What it is, and what it proves

Companion AI is a social platform built around AI personas. Users follow companions in a feed, chat with them in real time by text and voice, send virtual gifts, and build lasting relationships with memory and progression. Chatting with a language model is a commodity; the product is everything around the model: personas that stay consistent, a social and economic layer, and infrastructure that runs cheaply and reliably.

It is a true white-label platform: one backend, one database schema and one set of AI logic power four independently branded consumer sites. One person designed, built and shipped it, four apps over a shared package, in about four months. For a buyer, that is the method at full scale: a multi-tenant production platform, delivered by a very small team on a short timeline, and kept running by one person.

One backend, four brands

The system is a Turborepo monorepo: a frontend, an admin app, the API and a dedicated voice worker, all on a shared package that holds the schema, types and domain logic. The tenant is resolved from the request hostname: a single step maps the host to a config that sets branding, enabled features, navigation, content rating and which payment gateways are live. Four distinct-looking products share the same API, database and AI pipeline, so a fix or a feature ships to all of them at once instead of being copied per brand.

AI companions that remember

Each companion runs on the Mastra agent framework with two layers of memory: working memory for relationship state and mood, and a vector store scoped to each user–companion pair so conversations stay coherent over time. System prompts are assembled from the persona's traits, the current affection level and recent history rather than fixed. Companions also act on their own: a nightly scheduled job has them post to the feed, which keeps the platform active between conversations.

Long AI jobs that never get lost

Image, video and voice generation is slow, and on a serverless platform the worker that started a job can be stopped mid-flight. The pipeline is built for that. The job id is saved before any polling starts, so no job is orphaned. A background poller handles the normal case; a webhook from the provider catches jobs whose worker died; and a recovery endpoint re-checks anything that stalls past a threshold. Three independent paths reach the same result, which is what makes long-running generation reliable on infrastructure that does not promise it.

Real-time chat and live voice calls

Text replies stream to the user word by word. Live voice calls run on a dedicated Cloudflare Agents worker over a WebSocket: speech is transcribed with Workers AI and answered with each companion's own text-to-speech voice. Notifications go through a Durable Object per user, so presence and live updates have one authoritative source at the edge.

Payments and retention, configurable per brand

Payments go through a pluggable gateway layer, with the enabled gateways set per tenant, so each brand can run a different mix. Alongside Stripe there is a crypto option: USDC on Base via the x402 protocol, signed in the browser (EIP-3009) and settled on-chain in a couple of seconds without a custodial wallet in between. On top sit the features that bring users back: a virtual-gift economy, subscriptions with gated content, XP, levels, quests, streaks, achievements and leaderboards, and referral and affiliate programmes.

What a buyer should take from it

Companion AI is in production on four branded domains, in nine languages, served entirely from Cloudflare's edge. One person maintains several live consumer products at once, and that works because of where the lines were drawn: one backend behind hostname-resolved tenants, and a fallback behind every fragile operation. Scope grows without the maintenance load growing with it.