He interviewed 1,500 teams drowning in their own AI before he wrote a single feature. Then he built the floor they were missing.
Akash Sharma runs Vellum, and Vellum exists because of a feeling most AI builders know but few can name. He calls it quicksand. You write something that works on Tuesday, a new model drops on Wednesday, and by Thursday the ground under your application has shifted. Sharma's company sells the bedrock - a development platform where teams define, evaluate, and deploy AI systems without that sinking sensation.
Today Vellum is the connective tissue behind production AI at more than 150 companies. Redfin runs it across 14 markets. DeepScribe used it to cut iteration time. Drata, Swisscom, Cursor, Lovable, Salesforce and Headspace are on the customer list too. The pitch is unglamorous and exactly the point: take AI from quarters-to-production down to weeks, and do it without sacrificing quality. Sharma's phrase for the mission is test-driven development for AI - turning a craft that feels like guesswork into something engineers can measure.
What makes him an unusual founder for this corner of the industry is that he is not, by training, the engineer in the room. That job belongs to his two MIT co-founders. Sharma is the strategist who spent five years at McKinsey learning how large organizations actually adopt new tools - and then walked away from it to find out what happens when you build one yourself.
We've lived the AI development pain first-hand so our customers don't have to.- Akash Sharma, on why Vellum exists
The Vellum story doesn't begin at a hackathon or a dorm room. It begins at Dover, a Y Combinator startup where Sharma, Sidd Seethepalli and Noa Flaherty spent more than two years shipping production large-language-model features back when that was still an exotic thing to do. Seethepalli and Flaherty came out of MIT and DataRobot's MLOps team. Sharma came out of consulting. Between them they kept hitting the same wall: the tooling they wanted for building reliable LLM apps simply did not exist.
So in 2023 they built it, and joined Y Combinator's Winter batch to do it. The seed money - roughly $5M - arrived fast. But the more telling detail is what Sharma did before the product: he talked to over 1,500 people at every stage of running LLMs in production. That obsessive customer discovery became the company's foundation, and it shows in a roadmap that reads like a checklist of every real pain a working AI team feels.
Sharma describes himself as a connector - between academia, consulting, and the startup floor. It is a fitting self-portrait for someone whose superpower is translation: taking the messy frustrations of engineers and turning them into product, taking strategy-deck thinking and grounding it in shipped code. He is also, by reputation, generous with what he learns. Rather than hoard the playbook, he distills his findings back to the broader LLM community.
The result is a company headquartered on Madison Avenue rather than Sand Hill Road, betting that the future of AI belongs not to whoever has the flashiest demo, but to whoever makes the unglamorous middle - the evaluation, the versioning, the monitoring - actually work.
Generative AI has captured the imagination of nearly every industry, but turning that potential into reliable, production-ready systems remains a massive challenge.
Developing AI feels like writing software in quicksand - the ground keeps shifting.
We've lived the AI development pain first-hand so our customers don't have to.
We will be the best-in-class platform engineering teams around the world rely on.
Vellum's reach is the quiet proof of the thesis. From real estate to compliance to consumer wellness, these are teams that decided shipping trustworthy AI was worth a platform built for it.
Trained as an industrial engineer at Berkeley - an optimizer's brain wired before LLMs were a thing.
Five years at McKinsey's Silicon Valley office gave him a strategist's instinct before he became a founder.
Co-founders Sidd Seethepalli and Noa Flaherty are MIT engineers out of DataRobot's MLOps team.
His favorite way to describe building AI: quicksand. Constantly shifting ground under your feet.
Vellum is headquartered on Madison Avenue in New York, not the Bay Area.
Known for distilling his learnings back to the broader LLM community instead of hoarding them.
Make test-driven development the standard for building trustworthy AI - quarters to weeks.- The Vellum aspiration, in one breath