Skip to content
Sourabh DarojiMMMMMMMMMMMMMM
WorkExperimentsAbout
WorkExperimentsAbout

Let’s talk

Book a callEmailLinkedInXGitHub

FilterPixel · Startup · AI

FilterPixel, an AI-native platform that helps photographers pick their best shots

How might wehelp photographers find a wedding’s best shots in minutes, and still make every call?
By opening up the black box: every rating shows why
With an AI that only flags, and leaves the call to them
Through arrow keys that zoom to each face, no panning by hand
With Lightroom’s own keys, so there are no new habits
FilterPixel on an iMac, its panels resized so the photo takes the screenAnd by resizing the panels to give the photos their space

Case study

RoleProduct Designer
Timeline2020 – 2021
TeamPre-seed team of about eight
SkillsProduct Design, Interaction Design

More work

dots, pooling people’s Macs to run big AI models

Big models don’t fit on an ordinary Mac, so dots splits one across several. I designed and built the app, down to lending a Mac with a single link. We’ve pitched it to investors.

Independent · Product Designer & Design Engineer · 2025 – Ongoing
dots, pooling people’s Macs to run big AI models

NIVA, AI-native private cloud management

Private clouds are still run through screens from a decade ago. I’m designing an AI-native way to run them, and building it in code.

Nutanix · Design Engineer · 2025 – Ongoing
NIVA, AI-native private cloud management

IAM, managing identity and access at scale

The more consoles an enterprise runs, the harder it gets to manage identities and access policies across them. I’m designing one central console that makes it all easy to manage.

Nutanix · Lead Designer · 2025 – Ongoing
IAM, managing identity and access at scale

Tasks, the activity log of a private cloud’s control plane

Admins paged through hundreds of rows to find what failed. I led the redesign from research to ship, and made errors clear enough to fix without a support case.

Nutanix · Product Designer · 2023 – 2024
Tasks, the activity log of a private cloud’s control plane

Overview

FilterPixel is an AI-native desktop app that helps wedding photographers pick their best shots. A wedding photographer can spend upwards of 8 hours going through thousands of images to find the keepers. FilterPixel brought that down to about 30 minutes.

I joined in May 2020, when it was a pre-seed startup of about eight people, founded by Aayush Arora and Pratyush Goel. Thaxnay Kapdee and I were its two designers and shaped the product together; I also contributed to the code. In the ten months before I left for my master's, it went through three versions, each one changed by what working photographers told us.

Problem

We ran contextual interviews with nine professional photographers and asked about their workflow after a shoot. When we sorted the notes into an affinity map, the same three pains kept coming up:

  • Repetitive. Hours went to auxiliary tasks that added nothing to what they delivered.
  • Chaotic. Their workflow was held together with workarounds across fragmented tools.
  • Time-crunched. Client deadlines left no room for any of it.
Affinity diagram from user interview data showing clustered pain points
The affinity map from the interview notes

Constraints

  • Runway. A pre-seed company of about eight people, with some funding and not much time.
  • Fit in, do not disrupt. Photographers already had a workflow. FilterPixel had to slot into it and multiply their productivity, not ask them to work differently. It shows in the choices below: export to Lightroom, shortcuts they already knew, and panels that behave like the ones in their editing tools.

Three versions

One: the AI decides. Drag a folder in, let the AI filter and tag the photos, export to Lightroom. Usability testing showed the flaw: photographers did not trust an AI to make the call. Photography is subjective, and they wanted to stay in control.

UI screens from the first iteration of FilterPixel showing the initial tool-based approach
Version one: the AI filters and tags on its own

Two: the AI advises. We repositioned the AI from decision-maker to advisor. It flagged only the photos that were objectively bad, out of focus or eyes closed, and left the final call to the photographer. We also added a project hub, so all their work lived in one place. We gave it to insider photographers for a week, then ran contextual inquiry sessions. Two things came back: the interface wasted screen space, and the AI still felt like a black box.

UI screens from the second iteration showing the project management hub and AI advisor approach
Version two: a project hub, and an AI that only flags the obvious rejects

Three: the AI explains. Two changes. We made the panels resizable so the images could take the screen, the way they do in Lightroom and Photoshop. And every photo got feedback saying why it was rated as it was, with objective measures like focus quality and face detection scores.

The AI feedback panel showing per-image quality ratings and transparent reasoning
Version three: feedback on every photo, with the measures behind the rating
Panel-based customizable layouts with multimodal views for reviewing photos
Resizable panels and three views: grid, detail and faces

We also designed Faces View, a panel for inspecting faces without the constant zoom and pan. It became one of the product's most-loved features.

Dedicated Faces View panel showing extracted face details for quick inspection
Faces View: faces pulled out for inspection, without zooming in

What shipped

The final product shipped on a design system built from Apple's Big Sur UI kit, to match the Macs our photographers worked on, with keyboard shortcuts mapped to Lightroom and Photoshop so there was little to relearn.

The final shipped version of FilterPixel with the complete interface
The shipped version of FilterPixel
The FilterPixel design system based on Apple's Big Sur UI kit
The design system, built from the Big Sur UI kit
Keyboard shortcuts available in FilterPixel matching Lightroom and Photoshop conventions
Keyboard shortcuts, kept close to the photography apps they already used

Outcome

The problem. Culling that took 8 hours came down to about 30 minutes.

The product. Photographers told us the familiar shortcuts and layouts left little to relearn, and they kept calling out Faces View as intuitive.

The company. I left in March 2021. Six months later, FilterPixel was one of ten startups funded out of 4,785 pitches for 100X.VC's Class 05. Thousands of photographers took it up, and today the company counts more than 14,000.

What I learned

Trust in an AI is earned in steps. Version one asked photographers to trust it outright, and they refused. They accepted it once it was safe to ignore, flagging only the obvious rejects, and trusted it once it showed its reasons. People do not want an AI to replace their judgment; they want it to back their judgment. I met the same lesson again years later, on NIVA.

Being in the code closed the loop. I was in the research, the design and the build, so nothing got lost between understanding the problem and shipping the answer.