Winning Raw to Curated at GTC: A Local AI Workspace for Photographers

Studio Copilot won the Raw to Curated track at NVIDIA GTC's Hack to Create. It also got featured on NVIDIA Developer Live the following month. The product is simple: a local AI workspace for photographers and designers that handles the entire creative business workflow — raw photo curation, client review galleries, contracts, invoices, feedback collection — without ever uploading a single file to a cloud.

I want to talk about why "without ever uploading a single file" is the actual product, and why edge AI for creative professionals is the right architecture rather than a marketing line.

What photographers actually do

Most photographers I've worked with spend more time on admin than on creative work. The day breaks down something like this: shoot for two hours, sit in front of a screen for six. The six hours are split roughly evenly between three things: culling raw images down to a deliverable set, sending and managing client galleries, and the business layer (contracts, invoices, retainer renewals, feedback intake).

The culling is the bottleneck most professional photographers feel first. After a wedding, a portrait session, a real-estate shoot, you have between 500 and 3,000 raw images. Most of them are slight variations of each other — burst frames, exposure brackets, similar compositions. You want to keep the best version of each cluster and toss the rest, then deliver a curated set to the client.

Doing this by hand is brutal. A wedding photographer might spend a full workday culling a single event. Existing tools — Lightroom's stack feature, manual flagging — are unchanged in spirit since the 2000s. The work is essentially "look at every image, keep or reject, move on."

This is exactly the kind of task an AI vision model should be doing.

Why local

Cloud-based AI photo tools exist. They have a problem: they require uploading the photos.

For an amateur shooting their kid's birthday, that's fine. For a professional under contract, it's a deal-breaker. Wedding photographers sign NDAs with celebrity clients. Real-estate photographers shoot homes that aren't on the market yet. Editorial photographers work under embargoes. Sending those raw files to a third-party server — even one that promises not to look — is at best a contract conversation and at worst a contract violation.

The constraint flips when the AI runs on the photographer's own machine. No upload. No third-party. No log file in another company's data center. The model loads, the model runs, the model returns. The files never leave the laptop.

That changes which photographers can use the tool. It opens up the entire professional segment that the cloud version of this product can't address.

What we built

A desktop application running on a NVIDIA GB10 (the new Blackwell-based developer system NVIDIA announced at GTC). The core ML pipeline is OpenCLIP ViT-B/32 generating image embeddings, then visual-similarity clustering grouping similar shots. The UX wraps that pipeline in a culling interface that lets the photographer accept, reject, or override clusters in seconds per group instead of seconds per image.

The numbers from the demo: 1,400 raw images from a typical wedding session, clustered into 280 groups, culled by the photographer in under eight minutes. The same workflow done manually had taken her about six hours.

That's the core. Around it sits the rest of the workspace:

• Client review galleries — the photographer publishes a curated set, the client browses on their own machine (still no cloud), provides selections.
• Contracts — a small library of contract templates with fields auto-filled from the project metadata.
• Invoices — generated from the project, exported as PDF, ready to send.
• Feedback intake — a structured form clients fill in after delivery.

None of this is novel software individually. The novelty is that all of it runs locally, and the AI in the middle — the part that's actually doing work — runs locally too.

What the GTC judges noticed

Three things, in order of how often they came up:

1. The privacy story is real, not aspirational. The judges asked us to show them the network panel during the cull. We did. Zero outbound requests during the entire workflow. This is the part most "private AI" pitches can't deliver, and they could see it.

2. The hardware story matched the platform. The hack was built on NVIDIA's Hack to Create platform. The judges were specifically interested in projects that demonstrated what NVIDIA hardware enables. A GB10 running a meaningful workload locally — not a model card, an actual photographer-facing product — was exactly that.

3. The market is bigger than people think. The U.S. has somewhere around 200,000 working photographers, and the global number is several million. Almost all of them are running their own businesses. They are buyers of software when the software helps them. The total addressable market for a tool that saves a wedding photographer three hours a week is not small.

The pattern

Studio Copilot generalizes. The same architecture — local model, local UX, no upload — applies to a large set of creative-professional workflows. Editors. Designers. Architects with site photography. Lawyers reviewing case files. Doctors annotating scans. Anyone who has files they can't legally or contractually move to a cloud.

That market has been quietly waiting for hardware to be capable enough to run useful models locally. With the GB10 and the Jetson AGX line, we're past that threshold. The TAM for genuinely-local creative-AI tooling has just opened up.

If you're a photographer or designer reading this: the live build of Studio Copilot is in active development. We're onboarding pilots through aisoft.us. If you're the kind of business that needs creative-AI tooling that doesn't violate your client agreements, that's the door.

What I learned about hackathon strategy

A note for builders. The thing that won wasn't the model. OpenCLIP isn't novel. The clustering approach isn't novel. What won was taking a real workflow, real users, and a real constraint (locality) seriously, and shipping a product that respected all three.

Most hackathon entries pick the most impressive-sounding thing they can build in the time. The winners pick the most useful thing they can finish. There's a gap between the two and it's larger than you'd think.