All work

From a few phone photos to a product video, in under seven minutes, for about $0.30.

Avataar · Product Manager I, then AI Product Manager II · Velocity · 2023 to 2025

Velocity is a self-serve AI tool that makes product videos, cut-out images and lifestyle images for brands. I owned it end to end at Avataar, as the product manager on a team of nine. Avataar made product images and video with AI for online retail. HP was one of five or six enterprise clients on Velocity.

About $750
our compute cost for HP's first 2,500 videos, against a $1.25M agency quote
20,000+
videos made on Velocity in my time
About 25,000
signups in six months, with no paid marketing

HP's figures are one client's first phase, not a run rate. About $750 is what it cost us to make them, not what HP paid. Like about $0.30 a video, it counts compute only (the GPUs, the voice and the rendering), not our team's time, so it isn't like for like with an agency's price. Under seven minutes is the run alone. The costs, the studio times, the annotation time and the signup, setup and revenue figures are estimates.

How Velocity works, from a phone photo to a finished video.

The problem

An online store sells thousands of products, and a marketplace sells millions. Each one needs photos and a video, in every market it's sold in. Making them meant a photoshoot and then an editor. A studio charged $500 to $2,000 for one product video, and took about two to three weeks to make it. So most products never got a video at all.

Marketing wanted more than one video, and sooner. They wanted content before a product launched, and a few versions to test which one worked. Each new version took a long time to make. The brand and category managers who owned each product's content sent their feedback over email and spreadsheets. With thousands of products at a time, it got messy.

Why it was worth solving

In a three-month A/B test with an online furniture store, videos made on Velocity converted three to five times better than still images. The range spans product categories. The test covered more than 50,000 of its products. Both versions ran on the same products, at the same time, in the same region. The test counted purchases.

HP's first phase covered 28 markets and more than ten languages, and an agency quoted six months for it. Making the videos was the easy half. The markets were the harder half: each one needed a voiceover in its own language, and its own formats.

How we solved it

What I owned

My calls included templates over free generation, building Velocity into HP's own system, and cutting steps from signup. I was the product manager on a team of nine: three developers, a QA engineer, a product designer, three people on data annotation, and me. I reported to Sreejith Karumanaghat, our principal product manager, and from 2024 to our CPO, who was also a co-founder. I ran Velocity full time until mid-2024, then kept working on the HP contract until I left Avataar in October 2025.

I was on the client calls with HP and trained our own team.

A phone photo is enough

Let's say a brand is launching a new refrigerator. A marketer takes a few photos of it on a phone and adds the title, a description and the features. A product URL works as a starting point too, and so do videos and GIFs. One run gives back a video, three cut-out images on white and four lifestyle shots.

Velocity's input: a product URL pasted into a field, with an audience menu and checkboxes for AI voice and music beside a Create button.
The input as it looked in 2026, starting from a product URL, with an audience, AI voice and music.

Inside that run:

  • Velocity cuts the fridge out of its photo.
  • It scores a library of video templates against the fridge and picks the best match.
  • It places the fridge inside the box the template marks for it.
  • Stable Diffusion, an image model, makes the lifestyle images.
  • The script is the client's own translated copy for an enterprise, or AI drafts it for a self-serve user from the product's title, description, images and features.
  • ElevenLabs, a voice service, reads the voiceover in the market's language, and music goes under it.
  • It renders the video, and thousands of runs can go at once.
  • It goes out as a file to post anywhere, or straight into Meta Ads as an ad they could run.

A feature called the brand kit keeps all of it inside the brand's guidelines. We pushed videos into Meta Ads because a folder of finished files is still a job for someone.

Built for marketers, not editors

The whole platform was self-serve. Brand and marketing teams, agencies, freelancers and individual artists all used it. The brand and marketing teams weren't video editors, so it had to feel like Canva, not Blender. We left out the controls an editor expects. That capped how far a brand could customize a video, and it's why those teams could use it on their own.

They no longer waited on a photographer, an editor or another team. If they wanted a change, they swapped the voice, the music or an image right there. Because a phone photo was enough, they could start before a product launched.

A brand has to know what it'll get

Free generation, where a model makes each video from scratch, couldn't yet promise what a brand would get, so I chose templates.

We had more than 20,000 templates. They came from Creator, the video editor in the browser that I'd owned before, along with the engine that rendered them. Creator's users had made the templates by hand.

Velocity scored the library against each product, on more than 100 factors. Those included the category and subcategory (consumer electronics, then refrigerators), the number of cut-out and lifestyle images, and the brand's primary and secondary colors. The title, the description and the number of features counted too. The highest score picked the template. The weights for the images and the text were set in advance, not learned from data.

Choosing templates meant our three annotators spent about six months tagging every one by hand.

Avataar later built Varya, a video model, and it replaced the template approach. Varya isn't my work. I'd still start with templates.

Better models kept coming out

Our work was the platform, the templates and picking the right one, then rendering it well and fast. We didn't build the image or voice models. We plugged in the best on the market and kept our team on the platform.

Built into HP's own system

For HP, I decided to build Velocity into their own system, because the product data already lived there. That data stayed in sync with HP's other teams. A user entered a SKU ID (the product's code) and picked a language. Velocity pulled in the product's data by itself.

With one client, we worked out its data structure and sources, then connected it through an API built just for that client. Integrations weren't a feature yet. It was all plugged in within a week. I gave the demo myself, and the client signed the statement of work.

The impact

  • HP came back for a second phase.
  • Cutting steps from signup was my call, and we improved the UX alongside it. About 30% more new users finished setting up.
  • By our estimate, Velocity brought in more than $3M in enterprise revenue. I owned the product, not the deals.

Where it falls short

Looking back, I'd first let each client connect its own data sources to Velocity through an API. An enterprise could then turn its product pages into videos in no time. Enterprise and mid-market companies were the customers we were after, so that's where it would have counted most.

It struggled with chrome, glass and anything see-through. Cutting a product out of its photo is the hard step. On those surfaces the background shows through or reflects, so there's no clean edge to find. We'd kept editor controls out, so there was no tool to fix a cut-out by hand. HP's products in my time weren't glossy, so the limit didn't touch HP's videos.

Onboarding our second enterprise client meant rebuilding about 40% of the part that took in a client's catalog and data. We'd fit it to our first client, which wasn't HP. After that, we made our side of each integration plug and play, so the next client wouldn't need the same rebuild.

Velocity itself is still live, at velocity.avataar.ai.

Next case study

AN-Prez

Years before Facebook renamed itself Meta, I took a mixed-reality meeting from HoloLens to Quest.

Case study

Write to me about this one.

© 2026 Shubham Shrivastava