Teleprompter

Inspiration

To provide users with the best use case scenario of low-level reasoning models that can be accessed for free and completely privately

What it does

As of today, we do not really get history aware suggestions on the popular AI assistants available today, it is mostly generic and does not really capture the rabbit hole that we go through in researching a topic outside of that application.

Since the models offered by Chrome is completely local, we can track the user data for a recent period of time, index it and use that to provide useful current research context based suggestion for the user.

In addition to this we can also make use of the privacy functionality to redact prompts that are sent to these AI assistants and additionally provide redaction functionality to files too as the processing is completely done locally

How we built it

We build it using the crxjs/vite package which provides a good setup with vite and chrome extensions, additionally for styling standard tailwind css v4 was used.

Coupled with the the Chrome AI Prompt API and Summarizer API was used.

For data scraping from user visited sites, turndown a library that converts HTML to markdown was implemented

Challenges we ran into

The primary challenges working with the Model itself was to understand its reasoning and context limitations.

In general, Small Language Models with generic reasoning do not offer great opportunities to build useful workflows.

Another challenge was to capture user data that can actually be used to provide the suggestions and to build an async workflow to ensure that we get the right data.

Accomplishments that we're proud of

The identification of the use case, we think the redaction and bridging local user data would provide a great opportunity area in Generative UI as well as providing a better entry point in researching any topic using AI assistants via chrome.