Suchintan Singh spent years building ML infrastructure in environments he controlled. At Faire, he founded the ML infrastructure team and created a feature store for real-time search ranking. At Gopuff, he built the search engineering team. The platforms he created generated over $100M in GMV. Complex work, but the data was yours, the serving layer was yours, and the system behaved the way you designed it to behave.

Skyvern, the browser automation company Singh co-founded with Shuchang Zheng, was their third pivot. The second was a marketplace ML ranking platform, essentially productizing what Singh had already built twice. It had customers, $50K in ARR, $240K in sales commitments. They walked away from it toward a harder problem. Browser automation dropped him into territory his prior work hadn't mapped.

The customer calls tell the story. In a LinkedIn post describing 47 hours of prospect conversations in a single month, Singh reported that three out of four raised the same cluster of concerns: Can it handle 2FA? What about Cloudflare? The portal we need has CAPTCHA on every login. These aren't exotic targets. Lender systems. Government sites. Insurance carriers. Healthcare credentialing portals. The places where real business processes live, behind authentication walls built on the assumption that a human is sitting there. Five years ago, Singh notes, RPA bots ran on internal apps with hardcoded credentials. That world is gone.

Singh puts the engineering allocation bluntly: roughly 60% of the work in browser automation goes to anti-bot infrastructure. OTP forwarding. CAPTCHA vendors. Residential proxies. Fingerprint emulation. The agent reasoning, he says, is the smaller share of the effort.

The architecture reflects this. Skyvern's open-source repository contains the core agent logic. The anti-bot infrastructure exists only in the managed cloud product. The reasoning layer is open. The part that gets you through the door is proprietary. That split says something about where the value concentrates.

The technical choices point the same direction. Skyvern's agent doesn't parse the DOM to find a checkout button. It takes a screenshot and uses vision models to identify what looks like a checkout button. If the HTML changes between sessions but the page still looks the same to a person, the agent keeps working. Years of building ML ranking systems trained Singh to reach for model-based solutions. Here, that instinct met a problem it was actually suited for: sites where class names rotate, where form fields are coupled in ways the markup doesn't represent, where government portals go down at night or require you to call the IRS to proceed. The DOM lies, but a screenshot shows what a person would actually see.

One line from Singh's LinkedIn posts is worth sitting with:

Singh crossed from ML platforms to the open web and the environment stopped cooperating. The models improved. The web got harder. And the engineering hours followed the difficulty.