N0VA — AI Site Assistant Built to Ship

N0VA — AI Site Assistant Built to Ship

The studio's own AI assistant, grounded in real data, running under a dollar a month. The same engine we deploy for clients.

We built N0VA to answer for the studio. Ask it what we make, who we made it for, or whether we've done anything like the thing in your head, and it answers from real studio data, points you to the project that proves it, and stops short of pretending to know anything it wasn't built to know.

​​​​​​​

It runs on dopaminestudio.io right now. You can talk to it there.

Two things at once

N0VA is our own AI assistant, and it doubles as the working proof behind a service we sell. The same engine that runs it is the one we deploy for clients. So it's a tool that helps people move through our work, and a live demonstration that our AI Business Integration ships and runs in production instead of sitting in a pitch deck.

It runs on a RAG architecture built to hold a full working knowledge of the studio, at an operating cost under one dollar a month. The technical core is the one we built for EatAssistant. The voice and the knowledge base are its own. The name carries the toggle, straight out of the logo.

The space between rigid and reckless

Most site chatbots fail in one of two directions. The scripted kind only answers what someone anticipated in advance, so the first unexpected phrasing breaks it. The open kind, an off-the-shelf model dropped onto a page, answers anything, including things about your business that are wrong or invented. One is too rigid to be useful. The other is too loose to be trusted.

We built N0VA to sit exactly between them. Wide enough to hold a real conversation, narrow enough that everything it says traces back to something true about the studio. A guide that happens to sell, not a salesman that happens to answer.

Three decisions that shaped it

Each one was a deliberate constraint, and the constraints are the point.

Grounded in a fixed knowledge base.

N0VA retrieves from a closed set of studio content and answers from that. A raw model would have known more about the world and less about us, filling the gaps with confident invention. Tying it to a fixed base means it can only tell the truth we gave it. The limit is the feature.

A scoped system prompt as the second wall.

Retrieval decides what N0VA can draw on. The prompt decides how it behaves and where it stops. It holds the voice, the boundaries, and the instruction to decline cleanly when a question falls outside its purpose. The knowledge base keeps it factual, the prompt keeps it in lane.

Model sized to the job, not the hype.

N0VA runs on OpenAI's API, with the model and reasoning effort picked per task rather than defaulting to the biggest option. Most of what a site assistant does doesn't need a frontier model thinking hard. That's what keeps it under a dollar a month. The number isn't a compromise on quality. It's what happens when the scope is honest.

How it carries the brand

An assistant that talked like support software would quietly argue against everything the studio claims. So N0VA talks like Dopamine, turned warmer. It keeps the confidence, drops the edge, goes into detail when someone actually wants it. It has a personality of its own, held on the right side of the line where clarity wins over character.

The persuasion works by pointing rather than pushing. When something in the conversation lines up with a project, N0VA mentions it the way a person would, with a link. It reads intent before it suggests anything. The upsell is a recommendation from something that knows the catalogue, not a banner.

The honesty extends to the edges. When a question sits outside what N0VA was built for, it says so plainly rather than improvising a half-answer. That candor is partly character and partly economics. Spending tokens to guess at something off-topic serves nobody.

Visually, N0VA reads as native rather than bolted on. The toggle-O avatar, the acid-green accent against near-black, the type, all of it belongs to the site instead of sitting in the corner as a third-party widget.

The build

Retrieval over a fixed knowledge base. A closed corpus of studio content N0VA answers from and cannot stray beyond.

Scoped system prompt. Defines voice, boundaries, and refusal behavior. The second wall.

OpenAI API integration. Model and reasoning effort selected per task, holding operating cost under a dollar a month.

Contextual recommendation logic. Surfaces the relevant project when the conversation calls for it, with a direct link, and routes qualified intent toward contact.

Dopamine voice layer. A warmer register of the studio voice, with defined personality and clean refusal at the edge of its knowledge.

Native interface. Toggle-O avatar and the studio's color and type system, built to read as part of the site.

The same engine that runs this can run yours.

N0VA is what our AI work looks like when we build it for ourselves: scoped tight, grounded in real data, cheap to run, on-brand to the last word. Go talk to it, then tell us what you're building.