![Damani Brown](https://web.dev/images/authors/damanibrown.jpeg) Damani Brown [GitHub](https://github.com/Dbrown910) [LinkedIn](https://www.linkedin.com/in/getdamani) ![Melissa Mitchell](https://web.dev/images/authors/melissamitchell.jpeg) Melissa Mitchell ![Mari Viana](https://web.dev/images/authors/mariviana.png) Mari Viana [LinkedIn](https://www.linkedin.com/in/marianasviana) ![Thierno Thiam](https://web.dev/images/authors/thiernothiam.jpg) Thierno Thiam [X](https://twitter.com/localhost_droid) [GitHub](https://github.com/thierhost) [LinkedIn](https://www.linkedin.com/in/thierno-thiam-9753489b)

<br />


Published: October 14, 2025

<br />

![](https://developer.chrome.com/static/blog/ai-action-workshop-nyc/images/workshop.jpg)

In partnership with a group of startups in New York City, we hosted an "AI in
Action" technical workshop. The goal was simple: demonstrate how client-side AI
and built-in AI APIs can be directly integrated into their products for greater
speed, privacy, and a better user experience.

In total, we hosted 24 developers from 8 innovative startups. The morning
session featured talks from product and engineering experts, covering the
immense potential of open source, on-device models and the strategic importance
of bringing AI directly to the web platform. Technical specialists then provided
practical guidance to get started with the built-in AI APIs.

Attendees found the event to be highly valuable, discovering it was easy to use
these APIs regardless of their existing machine learning knowledge. They also
discovered new use cases for their applications through experimentation. We were
excited to validate our documentation. One group even helped us identify and
reproduce a complex bug, so we could submit a fix.

Our attendees successfully developed 10 prototypes using client-side AI. Take a
look at some of their projects and learn about their experience at the event.

## Adapter's event detection using Prompt API

[Adapter](https://www.adapter.com/)'s goal is to build "Mission Control for our
daily lives enabling maximum free-time not screen-time." They built a
proof-of-concept extension with two capabilities:

- Intelligent and aligned event detection which identifies events within a web page while browsing (concerts, restaurants, events) then adapts them with personal context including personal calendar availability, life preferences, and real-time social signals, all processed locally.
- Real-time content reranking that personalizes the ordering of content on a web page to your personal interests on demand.

Adapter used the Prompt API for parsing, reasoning, and calling tools, and the
Writer API for local text generation.
> "We're doing complex reasoning and multi-step workflows at the edge with
> limited privacy tradeoffs. This enables personal intelligence applications
> that were impractical when every inference meant shipping data across networks
> while deepening remote computation."
> -- Dillon Ponzo, Adapter's founding engineer

Chrome's built-in AI APIs enable lightweight models that process data locally,
preserving privacy by default. Adapter's founding engineers Aidan Crank and
Dillon Ponzo saw this as validation of their architectural philosophy.

"Most AI extensions capture context and send it elsewhere," explains Crank, who
previously worked on large scale ML engineering at AWS prior to joining Adapter.
"Chrome's built-in AI inverts that model. Your personal data stays on your
device, yet you get intelligent experiences previously unobtainable."

Adapter co-founder Adam Ghetti, whose previous company Ionic Security pioneered
autonomous data-centric security at scale, sees deeper implications: "Privacy
isn't a feature you add; it's an architecture you choose. Chrome's built-in AI
makes that architecture practical."

As Chrome expands these capabilities with multimodal support and richer APIs,
the edge isn't just catching up to the cloud. For personal and private AI, it's
about surpassing it where it makes sense.

## Sublayer categorizes images with the Prompt API to improve variations

[Sublayer](https://sublayer.com) created a fun app called Photoslider, designed
to revolutionize how you interact with and generate variations of your images.

To use PhotoSlider, the user uploads a photo, or captures a new one, and clicks
**Analyze**. The Prompt API sends these images to the model for analysis,
requesting the model identify 3 key attributes and give those attributes a value
of 1 to 10. For example, an attribute may be "traditional" or "luminosity."

On the frontend, the user is presented with the image and sliders to adjust the
values or introduce new attributes with custom values. Requests for changes and
the image are sent to server-side AI, so that a larger model with multimodal
output capabilities can generate a brand new image based on the changed
specifications This iterative process can be repeated endlessly, giving you
complete control until you achieve your ideal image.

[Scott Werner](https://www.linkedin.com/in/wernerscott/), CEO of Sublayer,
expressed his enthusiasm: "We've been genuinely interested in exploring what
local inference looks like. Having it readily available within the browser on
our users' machines by default is infinitely easier than relying on users to
navigate the complexities of installing models directly."

> [!NOTE]
> **Note:** Connect with the Sublayer team, [Scott Werner](https://www.linkedin.com/in/wernerscott/) and [Justin Pierce](https://www.linkedin.com/in/justpierce/). Subscribe to [Scott's newsletter](https://worksonmymachine.ai) for weekly experiments.

## Echo3D processes 3D models into rich searchable data

![](https://developer.chrome.com/static/blog/ai-action-workshop-nyc/images/echo3d-browse.jpg)

[Echo3D](https://www.echo3d.com/) provides a platform and API for enterprises to
seamlessly store, secure, and share 3D models and scans across diverse teams and
organizations. Their advanced 3D and text pipelines efficiently process
3D models into rich, searchable data.

At the event, echo3D developed 3D-to-text-to-3D tools that enhance the
visual comprehension of assets. They used the Prompt API to automate model
tagging, detect duplication, and streamline the cataloging and documentation
of extensive 3D libraries.

```json
{
  "description": "A 3D model of a large, tan-colored sandcastle with one main
    tower, four small towers, and staircases wrapping around. The towers
    have multiple windows. There is a main gate.",
  "tags": ["sand","castle","gate","tan","tower","staircase"]
}
```
[Video](https://www.youtube.com/watch?v=xGQ7mwFmWlY)

This gives every asset a rich, consistent, and searchable description, thus
making a user's entire library more organized and accessible. Furthermore, it
saves storage space and prevents versioning conflicts caused by redundant files.
> "Our team truly valued the event. We were particularly impressed by the ease
> with which AI features can be integrated into web applications using
> built-in AI. The ability to prototype and deploy secure, client-side AI
> features in such a short timeframe was a significant advantage."
> -- [Alon Grinshpoon](https://www.linkedin.com/in/alongrinshpoon/), CEO at [echo3D](http://echo3D.com)

> [!NOTE]
> **Note:** You can connect with the participants of the echo3D team: [Alon Grishpoon](https://www.linkedin.com/in/alongrinshpoon/), [Lea Broudo](https://www.linkedin.com/in/lea-broudo/), [Alex Wysoczanski](https://www.linkedin.com/in/alex-wysoczanski/), [Olivia Huang](https://www.linkedin.com/in/olivia-huang-21b73a14b/) , and [Andres Product](https://www.linkedin.com/in/andresproduct/).

## Spot2 created structured metadata from uploaded images

[Spot2](https://spot2.mx/)'s application lists real estate properties for rent
in Mexico City. Often when these properties are listed, critical metadata and
other valuable information is missing from the listing. The details are highly
variable, depending on the supplier's attention to detail. Spot2 spent the day
enhancing data quality with the Prompt API.

Their team developed a feature to automatically process and organize how listing
metadata is structured. This process occurs the moment a photo is uploaded,
which helps with completeness and consistency. By performing this task
client-side instead of server-side, the feature is cost-effective.

If launched, this feature promises a dual impact: a marked improvement in data
quality and a noticeable reduction in operational expenses. "We anticipate that
higher quality listings will also translate into an increased conversion rate."

## Join us next time

- [**Register for the next AI in Action Workshop**](https://luma.com/or5gf23q) on November 7, 2025 in San Francisco.
- [**Join the Built-in AI Challenge 2025**](http://goo.gle/ChromeAIChallenge2025). We're hosting a virtual hackathon for all developers. Create web applications or Chrome Extensions using built-in AI APIs, for a chance to win one of $70,000 in prizes.