Dhristi View on GitHub

SIH26171 · On-device visual perception

AI that browses for you.
Your details stay with you.

Dhristi lets a browser agent understand what is on your screen without sending your screen anywhere. Faces, personal details and form values stay on your device. The agent only sees the layout it needs to help.

Watch the 25-second demo How it works
The Himalayas and the Indian subcontinent seen from the International Space Station.
The Himalayas from the ISS. Photo: NASA (iss074e0002581).
  • 0personal values in 48 requests to a real cloud planner (Groq) across five end-to-end runs
  • 305 msfor the local prototype to capture a page, find the face and protect it, in one recorded run
  • CPUis all the redaction model needed in our tests. No graphics card.

Measured on synthetic test pages, 29 September 2026. Evidence: Benchmarks/results/e2e-v5-boundary-v02.json.

Agents need to see.
You need privacy.

AI browser agents can fill forms and move through websites for you. To do that, most of them send a screenshot of your screen to a model in the cloud.

That screenshot holds whatever you had open: your name, your email, account numbers, photos. Dhristi keeps those on your computer and gives the agent a protected outline of the page instead.

How it works

Four steps. Your details never leave.

  1. 01

    See

    Dhristi reads the page and its controls right in your browser.

  2. 02

    Protect

    A model on your own computer finds names, emails, numbers and faces, and swaps each one for a placeholder.

  3. 03

    Plan

    A larger AI model in the cloud plans the next step. It only ever sees placeholders like EMAIL#1.

  4. 04

    Act

    Your browser checks the plan, puts your real details back in locally, and asks you first before anything risky.

A synthetic service desk page showing a photo, a name, an email address, an account number and a password field.
What you see The full page, on your screen only.
The same page after protection: the photo, email and account number are pixelated and the buttons are marked as controls.
What stays protected The face, email and account number are masked on your device.

Screens from the working prototype, using a synthetic demo account.

See it working

The prototype, recorded live.

25 seconds, recorded from the local prototype with synthetic demo data. It captures a page, masks it on the device and shows the only data that may leave.

In our end-to-end tests the real cloud planner received EMAIL#1 and PERSONNAME#1. The real email and name were typed into the page by the browser, on the device.

The privacy boundary

What stays. What leaves.

Stays on your device

  • Screenshots and pixels
  • Names, emails and field values
  • Photos and faces
  • Web addresses and free text

Sent to the cloud planner

  • Which buttons and fields exist
  • Where they sit on the page
  • Placeholders such as EMAIL#1 instead of real values

A smaller view is not a guarantee of anonymity: the layout of a page can still reveal something about it.

The limestone shoals between India and Sri Lanka seen from orbit.
The shoals between India and Sri Lanka. Photo: NASA (iss071e700080).

Smart India Hackathon 2026

Built for SIH26171.

ISRO's problem statement asks for on-device visual perception for lightweight browser agents. Dhristi is our answer: seeing the page and removing personal details happen on your device, and only the protected result goes to a cloud model for planning.

Dhristi is a student engineering prototype. It is not a released product.

Read the problem statement

RV University · B.Tech CSE

The team.

  • Gopreet

    API and persistence

  • Hiranmayi

    UX and accessibility

  • Varun

    Core implementation

  • Koushaik

    Testing and reliability

  • Francis

    Integration and pitch

  • Niharika

    Research and narrative

See it work for yourself.

Watch the 25-second demo Explore the source