In the spring of 2024, Yuval Steuer sat in the front seat of the car he had been living out of for most of the year and opened the landing page for his company. In the top-left corner, the page read “Kaps.” Twelve months earlier, the company did not exist. Now, tens of thousands of video professionals across the world are using it to automate one of the least glamorous parts of post-production: subtitles.
Traditionally, the workflow to transcribe videos would start with an editor who would extract the audio from a video, send it through OpenAI’s Whisper transcription model, receive words back with timing information attached, split the transcript into subtitle chunks that Adobe Premiere Pro could understand, wrap the output in the SubRip text format, one of the most common subtitle-file formats, and hand the file back to the editing timeline.
When any part of the workflow fails, the failure is often invisible. The transcript comes back with words but no reliable timestamps. The Hebrew renders left to right instead of right to left, and the editor sees a string of characters that begins with the period and ends with the first word of the sentence. Each failure mode, taken individually, seems minor. But together, they help explain why studios can spend three hours typing subtitles by hand for every twenty minutes of video.
“We saw editors typing subtitles manually, and it just felt insane,” Steuer said. “Like, this is exactly the kind of thing computers should be doing.”
How Kaps Turned OpenAI’s Whisper Into a $1 Million AI Business
Steuer and his co-founder, Roey Lalazar, built a product designed to solve those workflow failures. They called it Kaps. Twelve months after launch, Kaps had reached more than $1 million in annualized recurring revenue, more than 40,000 users, and $70,000 in monthly recurring revenue, without taking a dollar of outside venture capital. The company’s growth reflects a less-discussed side of the AI boom: small teams identifying narrow workflows that foundation models can absorb, then carefully building the surrounding product infrastructure for the people who already have the problem.
Although Whisper attracted significant attention after OpenAI open-sourced the model in September 2022, the model itself was not a finished product. Whisper produces a stream of words, sometimes with timestamps, in any of roughly ninety-nine languages. A video editor cannot simply drop raw transcription output into a Premiere Pro timeline. Instead, the editor needs captions that begin at the correct frame, end at the correct frame, break at the natural pause, render in the writing direction the language uses, and arrive in a format Premiere can import without modification. Steuer and Lalazar focused on closing the gap between Whisper’s raw output and a production-ready subtitle file.
Adobe Premiere Pro’s native transcription feature, when Steuer began the work, supported a list of major Western and East Asian languages, but not Hebrew. Whisper did.
“The whole thing started because Adobe didn’t support Hebrew and Whisper did,” Steuer said. “That was the opening.”
Supporting a right-to-left language required the system to handle bidirectional text rendering, directional shifts that occur around numbers and proper nouns, and segment-boundary detection that recognized Hebrew’s prosodic patterns rather than the punctuation conventions of English.
The Workflow Gap Between AI Models and Real-World Video Production
In summary form, the pipeline Kaps shipped looked like a series of unglamorous engineering decisions. Extract the audio in the format Whisper expects. Send it through Whisper at the appropriate model size. Hold on to the word-level timestamps the model returns. Detect the prosodic and syntactic boundaries that produce natural subtitle breaks rather than mid-clause cuts. Encode the result as SRT. For Hebrew, apply the right-to-left treatment Adobe’s import path requires. Hand the SRT back in a form Premiere can ingest in two clicks.
The initial market was Hebrew-speaking users. From there, Kaps spread until it supported Polish, Hungarian, Arabic, and Portuguese, and the work to add a new language could be lightweight enough that it lived inside a single configuration change.
“Once it worked in Hebrew, it was obvious this wasn’t just an Israeli problem,” Steuer said. “Everyone hates doing subtitles manually.”
Once the product started working, it was time to tackle distribution. A well-known video-editing influencer, who knew Steuer, posted about Kaps on Instagram. The audience consisted almost entirely of video editors who had the problem the product solved, and within thirty minutes, Kaps received $4,000 in pre-orders. “That first influencer post was crazy,” Steuer said. “We made four thousand dollars in pre-orders in half an hour, and that was the moment we knew this was real.”
The post became the company’s strategy. Kaps eventually worked with seventy to eighty influencers across markets. Driving the company’s expansion has been exceptionally young Head of Growth Avi Srivastava, who spearheaded the influencer and advertising strategy across Kaps’s international markets. Avi helped scale Kaps beyond its initial Hebrew-speaking user base, transforming a specific initial solution into a global platform used by over 40,000 video editors. “Avi has been instrumental in Kaps’ global growth”, co-founder Lalazar says.
The company’s hiring strategy followed the same logic as the engineering department. Use the fastest and simplest tool that solves the problem, then engineer carefully around its limitations. Kaps hired specialists by language and function through OnlineJobs.ph, the Philippine remote-work marketplace, including copywriters, plugin developers, lead-generation analysts, and outreach operators. The company built individual landing pages for each market on a shared Tailwind UI Salient design template, with copy rewritten for each audience. Revenue paid the bills, which forced a kind of clarity.
“Bootstrapping is brutal, but it’s very clarifying,” Steuer said. “You can’t lie to yourself for long.”
Why Workflow Engineering, Not AI Models, Created the Business Opportunity
One less-publicized part of the company’s early history was the year Steuer spent living out of his car while Kaps found its footing. He documented the experience on X, where his account has become relatively well-known within indie-founder accounts in the Israeli AI community. “The car thing sounds dramatic now,” he said. “But at the time, it was just math. I wanted the lowest possible burn so I could keep building.”
Kaps reflects a category of AI business that much of the mainstream AI conversation has treated as secondary. The loudest coverage has gone to the labs releasing frontier models. But across many professional workflows, the commercially valuable work has often emerged one layer above the model itself: extracting audio, preserving timestamps, handling language-specific formatting, integrating with the tools professionals already use, and packaging the result into something a working editor will pay for.
Kaps began as a solution to Adobe Premiere Pro’s lack of Hebrew transcription support – a year later, it had become a bootstrapped software company with more than $1 million in annualized recurring revenue. The models enabled the product, but the workflow engineering made it a business.
