The Build with Gemini XPRIZE winner treats 3D models as editable programs, then uses AI agents to help produce and check them.

At a Glance

XPRIZE reported that Polyfork’s AI-agent-supported catalogue was publishing more than 175 models a week when the award was announced.

Imagine asking for a taller traffic cone that has another reflective band, and still looks right beside the other objects in your scene. A regular 3D model may let you stretch it, but that can distort the shape. Properly changing it might mean opening specialist software and editing the model yourself.

Polyfork takes a different approach: its assets are designed as small programs that can be adjusted and rebuilt. The idea helped the project win the $500,000 first-place prize in the Build with Gemini XPRIZE, a competition challenging builders to launch AI-powered businesses.

The problem is not finding a model

Developers can find plenty of 3D models online. The harder part can be finding one that fits: the right dimensions, level of detail, style, and technical format for a particular project.

That is the problem Polyfork founder Lucas Martinic describes in the project’s competition submission. A fixed model may be easy to download but difficult to adapt. Stretching it can make it look wrong; editing it properly takes time and 3D skills. And even when individual assets work, they may not look like they belong in the same scene.

Polyfork is aimed at web 3D developers, game and XR creators, and teams building tools that assemble interactive scenes. Its premise is straightforward: make assets easier to reshape, and make related assets feel like a set.

A model with “knobs”

Polyfork assets can expose adjustable parameters—think of them as knobs for properties such as colour, proportions, dimensions, or repeated parts. Change a setting and the model’s geometry can be rebuilt around it.

The traffic-cone example: In the project submission, one version has two reflective bands and 280 triangles. At another setting, it gains a third band and has 472 triangles. The model has been rebuilt to accommodate the change; it has not simply been stretched.

Assets can also be grouped into themed kits, designed around shared rules such as colour palette and scale. That is intended to make it easier to place several pieces together without every model feeling as though it came from a different world.

The code-based format matters, too. A developer—or an AI tool—can potentially inspect and adjust named parameters more directly than it can with a fixed, opaque model file. Polyfork also describes an API and MCP server for letting agents search for and retrieve assets.

• The core distinction: Polyfork is not just about making more 3D models. It is about making models that can be adjusted and used by software.

Inside the AI production line

Polyfork’s competition submission describes a workflow that combines AI agents, automated checks, and human review:

1. Martinic provides a kit concept.

2. An agent helps turn it into a production brief, while Gemini generates a visual reference.

3. Another agent writes code to build the geometry.

4. A rendering and measurement harness checks the result.

5. A separate agent reviews it against the brief and kit rules.

6. A person can review the output and request revisions.

This is important because “the model is finished” and “the model is correct” are not the same thing. Martinic’s submission says the judge agent sent back roughly two-thirds of the attempts it reviewed, often because of proportion problems or a mismatch with the brief. That is a project-reported figure, not an independently audited metric.

The submission says text generation ran through Vertex AI on Google Cloud, with Google AI Studio as a fallback. But it would be misleading to say that Gemini alone creates every finished asset: the described pipeline also uses coding agents, automated checks, a separate reviewer, and human judgment.

Early traction—and what it doesn’t prove

XPRIZE reported that Polyfork had 583 models, 54 paying customers, and $4,428 in first-month revenue. It also said the AI-agent factory was publishing more than 175 models each week. These numbers show that Polyfork had a growing product and paying customers during the competition.

They do not, by themselves, tell us whether customers returned, whether revenue was recurring, or whether the company’s costs were sustainable. The public announcement does not break down the $4,428 by subscription, lifetime purchase, or other transaction type, so it would be inaccurate to assume all 54 customers were subscribers.

Polyfork’s current pricing page lists a free tier and a Pro plan at $15 per month. Its separate platform plans are aimed at businesses that want to serve assets to their own users. These are current website details, not necessarily the prices or sales mix during the competition, and should be checked again before publication.

What builders can take from Polyfork?

• Pick a specific pain point. Polyfork focuses on adapting and integrating 3D assets, not automating every part of 3D creation.

• Make the output useful to software. An asset with named, editable parameters can be easier for developers and agents to work with.

• Check what the AI actually produced. The project’s described rendering and measurement harness is a reminder to verify results instead of trusting an agent’s own account of its work.

• Keep people involved where judgment matters. The founder’s submission describes human review as part of the quality process.

• Treat early revenue as evidence, not a verdict. Paying customers are encouraging, but retention, repeat use, costs, and margins are also needed to assess a business.

The bigger bet

Polyfork’s larger bet is that AI-assisted 3D creation will need more than tools that generate individual objects. It may also need collections of assets that are coherent, customizable, and easy for both people and software agents to use. The XPRIZE award and disclosed early revenue make Polyfork a notable experiment—but the public figures do not yet establish long-term retention or profitability.

FAQs

1. What is Polyfork?

A library of customizable 3D assets designed for developers and AI-assisted workflows.

2. Who founded it?

Lucas Martinic, whom XPRIZE lists as Polyfork’s founder.

3. How does Polyfork use Gemini?

Its submission describes Gemini supporting parts of a workflow that also includes agents, automated checks, and human review.

4. How much revenue did Polyfork report?

XPRIZE reported $4,428 in first-month revenue from 54 paying customers.

5. Did Polyfork win the competition?

Yes. It won first place and the $500,000 grand prize.