2018 – 2021
Vectorly
Four products. Three years. One acquisition. The story of how a video compression startup found product-market fit on the fourth try — and what happened after.
From dot Learn to Vectorly
Vectorly was born from the pivot of dot Learn, an e-learning app I had built to serve students in West Africa. When the consumer product failed to scale, we kept the underlying video compression technology — a technique I'd developed to compress educational video by 10x — and tried to commercialize it directly as a B2B company.
We raised roughly $600k in pre-seed funding from experienced investors including founders of Hulu and Elemental Technologies, giving us about 2.5 years of runway to figure it out.
What followed was three years of hard work and four distinct products before we found something that actually worked.
Product 1: "Vimeo for Emerging Markets" (2019)
The first B2B product was a video hosting platform — like Vimeo, but with our compression built in, optimized for low-bandwidth markets. The pitch: edtech companies in India, Nigeria, and Southeast Asia could host their content with us and their users would get a 10x better experience on mobile data.
After three months of cold-emailing 200 companies, we landed 2 paying customers at $50/month. They were real, they were using it — but the math was brutal. We'd need 2,000 similar customers to raise a seed round, and there simply weren't that many edtech companies at the right scale in these markets.
We convened a board meeting at a WeWork in San Francisco in November 2019 — advisors included founders of Hulu and Elemental — and unanimously voted to abandon the edtech hosting product and go full-force into video vectorization for media and entertainment.
Product 2: Video Vectorization (late 2019 – 2020)
The core technology — converting video into a vector format — was genuinely novel. We spent December 2019 in a coding sprint in Bangalore, and by early 2020 had our first real vectorized video demo. A Simpsons clip. It looked great.
We presented at Berkeley SkyDeck Demo Day in February 2020 in front of ~1,000 investors, then spent three weeks on follow-up calls with over 100 VCs. In late February — with early news of a virus called Coronavirus starting to circulate — we took a $150k check from an investor on reasonable terms. That decision turned out to matter a lot more than we knew at the time.
From April through August 2020 we improved the technology, hired PhD engineers in computer vision, and prepared demos. Then we went out to pitch the biggest players: Netflix, YouTube, Disney, Warner Media, Crunchyroll. We had meetings with the right people at all of them.
The feedback was consistent: technically impressive, but not a burning problem. Saving a large studio a few million per year wasn't compelling enough to justify the integration complexity. And our vectorization only worked well for flat 2D animation — a small fraction of their content.
In October 2020, we gave a talk at Demuxed, the video engineering conference, to the entire industry. The reaction was the same: cool technology, not a priority.
Two years after pivoting to Vectorly, we had built and validated the technology — and invalidated the market thesis in two months of customer conversations.
Product 3: AI Upscaling SDK (December 2020 – mid-2021)
It was November 2020. We had no product, no revenue, and less than a year of runway. I gave myself one month to figure out what to do next.
One idea kept surfacing: instead of replacing a video codec, use AI to upscale low-resolution video back to high quality. Compress the video to a tiny size, then use a neural network running in the browser to restore it. The end result: better quality at a fraction of the data cost, with no changes needed to backend infrastructure.
I had no background in neural networks. I spent December 2020 doing a crash course — learning enough about convolutional neural networks to reverse-engineer PyTorch models in pure JavaScript, then rebuilding them layer by layer in WebGL. I worked through Christmas. By New Year's, the WebGL model was producing results pixel-for-pixel identical to the PyTorch original.
By late January 2021, working with Yuvraj, a computer vision engineer on the team who had also taught himself WebGL in two weeks, we had the world's first practical real-time AI video upscaling library running in the browser — faster and higher quality than anything in the open-source ecosystem.
We posted on Reddit, HackerNews, and video engineering forums. We got 30 serious signups from qualified companies. We followed up with every one of them.
By May 2021, none of them had put us in production.
Product 4: Virtual Backgrounds SDK (June 2021 – acquisition)
I did one more round of market analysis in May–June 2021, going through every plausible application for real-time AI video processing in the browser: satellite imagery, police body cams, telemedicine, sports broadcasting. Nothing clicked.
Then I looked back at the signups we had gotten for the upscaling SDK. About half of them were video conferencing companies. When we talked to those companies more deeply, virtual backgrounds kept coming up on their product roadmaps.
We spent June 2021 rebuilding the SDK around virtual background segmentation. Two days after updating the website — before we'd even sent a single outbound email — a company signed up and asked how to pay. We hadn't set up billing yet. Within 24 hours, we got on a call, integrated their product, set up Stripe, and received our first real customer payment since dot Learn had shut down in 2018.
By August 1st: three paying customers (two on enterprise contracts), twelve companies actively testing, and Hopin in production. By September, after releasing an ultra-efficient WebGL version, Hopin and dozens of other large companies had committed to pay or were live in production. We went from $0 to $20k MRR in the first few months.
Three years after raising our pre-seed, we finally had commercial validation.
The Hopin Acquisition (December 2021)
Hopin — then one of the fastest-growing companies in history — had become our largest customer. At some point during the commercial traction phase, they approached us about an acquisition.
I didn't take it lightly. We had 193 investors lined up for a seed round. We had real momentum. I built a full decision-tree analysis — acquisition vs. fundraising, every path and its probability of success. I thought hard about what it would mean for our investors, our team, our customers.
In the end, we chose to join Hopin. The opportunity to work on AI features at scale — Hopin's StreamYard product had millions of users — was compelling. And personally: I like building, and I dislike admin, sales, fundraising, and legal. Joining Hopin meant I could focus on engineering.
The Vectorly team joined Hopin on December 1, 2021. The AI SDK was integrated into StreamYard as features like AI Touch-Up, Background Noise Removal, and AI Clips, each used by roughly 100,000 monthly active users.
The End of Hopin (2022 – 2024)
After the pandemic, Hopin's core business declined sharply. The company went through three rounds of layoffs between 2022 and 2023, shrinking from 2,000 employees back to around 300. By 2023, the focus had narrowed almost entirely to StreamYard.
In April 2024, the investors voted to sell the company to Bending Spoons, a private equity firm. Bending Spoons promptly laid off most of the remaining staff — including the entire Vectorly team.
The team dispersed. I started Katana.
What Became of the Technology
The AI upscaling work didn't disappear. I open-sourced the core library as WebSR, and built a free consumer tool on top of it — Free AI Video Upscaler — as a side project. It grew to 250,000 monthly active users without any marketing, and a case study on web.dev documents how it works.
The Video Vectorization patent was eventually open-sourced. You can watch the Demuxed talk for a full technical walkthrough.
Press & Recognition
Vectorly started as dot Learn, an e-learning app for students in West Africa. Read that story too.