Lumen Atelier

This one is a different kind of entry in this collection. Every other case study here is something I built, with an AI co-pilot alongside me. Lumen Atelier is the inversion: I set the constraints and played creative director, and the studio itself was conceived, architected, and implemented by Claude (Fable 5) as an open-ended commission. The brief was deliberately short — make it visually dazzling, run it in Docker, and use AI in a way that hasn't been done to death — and then I got out of the way and shaped the result through direction rather than code. The MIT license is co-signed for exactly that reason.

It belongs on this site because the thing that came back is genuinely on-thesis. The whole portfolio argues that the interesting engineering in AI lives in the loop around the model — the perception, the judgment, the gates. Lumen Atelier is that argument turned into an art studio: a system that doesn't just generate, it looks, critiques, and revises.

The core idea is a genuine perception-action cycle. The studio renders its own work headlessly, captures real frames from the running piece, and shows those frames to a vision model acting as a gallery critic. Draft, look, critique, revise — and only the Critic can admit a piece to the permanent collection. Nothing reaches the gallery on the strength of a prompt alone; it has to survive being looked at.

The entire creative argument is preserved and browsable for every piece: the brief, each draft (including the rejected ones, still rendering live), the compile errors, the Critic's scores, and the notes that sent a draft back. One of the first works — Alchemy of the Lattice — was approved only on its third draft, after the Critic twice rejected an earlier version as “oversaturated lime-green plasma.” You can read that whole exchange on the piece's page.

Three roles, three models, each cast to fit the job. The Muse (Haiku) writes the concept brief — one idea, a disciplined palette, concrete motion — and can ground it in real art-history research. The Artisan (Sonnet) realizes the brief as a GLSL ES 3.00 fragment shader in the demoscene tradition, streaming its work token-by-token onto the live studio floor and repairing its own compile errors. The Critic (Opus) is the gate: it examines four rendered frames spanning the animation, scores composition, color, motion, and fidelity to the brief, and issues a verdict — approve, revise with concrete notes, or, on the final round, decline.

The casting is just configuration — the whole ensemble can be recast through environment variables — but the division of labor is the point: a cheap, fast idea-generator, a mid-tier craftsman that does the volume of work, and the most capable model reserved for the one job that needs real judgment.

The medium is the demoscene's: full-screen fragment shaders, pure math, no textures or assets. That choice has a quietly radical consequence — every finished piece ships as text, and renders live, in real time, on every visitor's GPU. The gallery is never a video or a saved image. It is the artwork itself, executing in the browser. The same shader source that the Critic judged in the headless renderer is what runs on your screen.

Because visitors execute unvetted, machine-written shaders, the gallery's WebGL layer had to be hardened like production code: per-canvas frame-time watchdogs, automatic recovery from lost graphics contexts, and visibility-gated rendering. It's a nice illustration of the difference between a demo and a system — the interesting work was making machine-generated art safe to run on a stranger's machine.

Even in an art studio, the co-pilot-versus-autopilot line shows up — this time drawn around cost, not production risk. The studio self-commissions and iterates autonomously all day: research, ideation, draft/critique cycles, with no human in the loop, because it's spending its own budget on its own ideas. But visitor commissions spend real compute on someone else's idea, so they pass through a curator.

Anyone can browse with no account. Proposing a commission requires Sign in with Google, and every proposal lands on the curator's desk and consumes zero tokens until a human explicitly approves it — declined proposals never run, with per-user rate limits on top. The curator can also send any piece back to the studio with notes, fork it into a fresh redraft, or pull it from the collection, and every piece carries a usage ledger of the actual model spend that made it. The autonomy is real, but the gate on consequential spend is human — the same instinct as every other project here.

It runs continuously in production on Railway — three services, self-commissioning a new piece on a fixed cadence — so the gallery is a living thing, not a one-time render. As the director rather than the author, my job was taste and constraint: deciding what “good” meant, where the gates belonged, and when a result was worth keeping.

The reusable idea here isn't the art; it's the loop. A generator that can perceive its own output and a critic with the authority to reject it will beat fire-and-forget every time — in art, and in the far less whimsical systems the rest of this site is about. That's why this piece sits alongside the others: same conviction, different canvas. It's open-source under MIT if you'd like to read how it fits together, and the live gallery is the best way to see it.