There is a specific kind of dread that lives in software. It arrives at 3:07 in the morning, in the form of a phone that will not stop buzzing. Something broke in production. A graph is red. A customer is angry. And somewhere, an engineer is fumbling for a laptop in the dark. Viraat Das has been that engineer. He would rather nobody ever had to be again.
That instinct - remove the pain, remove the waste, remove the wait - runs through everything Das has built. He is the cofounder and CEO of Exla, a two-person startup that went through Y Combinator's Winter 2025 batch. The company is small enough to fit around one table and ambitious enough to take on one of the least glamorous, most universally hated corners of engineering: keeping production alive.
01 / The originTwo and a half years, then straight to work
Das studied computer science at Purdue University and finished in roughly two and a half years - a compression of the usual four that tells you something about how he treats time. He didn't linger. He shipped early into the working world, joining Amazon as a machine learning engineer, where he worked on personalized search and on the unglamorous but essential infrastructure that makes large models run faster and cheaper.
That Amazon chapter matters more than it might seem. Personalized search is a problem of scale and latency - millions of queries, milliseconds to answer, models that have to be squeezed until they fit. It is optimization work. And optimization, it turns out, is the lens Das keeps returning to.
02 / The partnerA company built by two people who study sleep
Das didn't build alone. His cofounder and CTO is Pranav Nair, whom he met at Purdue. Before Exla, Nair was an operating-system engineer at Apple, where he led the sleep and hibernation features that run across Apple's devices - the quiet machinery that decides what a computer does while you are not looking.
There is a nice symmetry there. One founder spent years on how machines rest; the other is now building software that refuses to. Exla is a company obsessed with what happens after everyone has gone to bed - the deploy that quietly broke, the log line nobody read, the regression that hid for hours. Two Purdue engineers, one from Amazon and one from Apple, betting that production systems should watch and heal themselves.
03 / The first productShrinking giants down to the edge
Exla's earliest public work aimed at a hard problem in AI: the biggest, most capable models are also the heaviest, and most of the world's hardware is small. Robots, cameras, sensors, an NVIDIA Jetson bolted to a drone - none of them has a data center's worth of memory. So Das and Nair built an SDK that aggressively optimizes transformer and vision models to run on exactly that kind of constrained hardware.
The numbers they published are the kind that make hardware engineers lean in: memory footprint cut by up to 80 percent, inference sped up by 3 to 20 times. In plain terms, a model that used to demand a beefy GPU could suddenly run on something you can hold in one hand.
Reported figures from Exla's SDK for running transformer and vision models on constrained edge hardware.
04 / The turnFrom smaller models to self-healing systems
Startups evolve, and Exla pointed its optimization instinct at a new target: the production system itself. The result is Libra, an autonomous AI site reliability engineer - the kind of role companies usually staff with a rotation of humans and a lot of coffee.
Libra is designed to do the whole loop. It learns what normal looks like for a system, then watches the logs and traces for the moment things drift. When something breaks, it correlates the anomaly against recent deploys, works out which change caused the failure, and opens a fix pull request - regression test included. It lives in Slack, sends proactive alerts with an estimate of customer impact, and can be queried over MCP. The point is not to help the on-call engineer. The point is to be the on-call engineer.
05 / The builderA workshop that never closes
To understand how Das works, look at his public code. His GitHub is less a portfolio than a workshop with the lights permanently on - hundreds of repositories spanning Swift, Rust, and TypeScript. There's a meeting recorder built specifically for AI agents. A PDF viewer and editor. A Mac window switcher. A meta-tool for orchestrating hundreds of coding agents at once. And, sitting quietly among the utilities, a project called Humanity Archived, an effort to collect and preserve the stories of people.
It's a telling mix. Most of the repos are sharp, practical, made-because-I-needed-it tools. Then there's the one that reaches for something larger. That range - relentlessly useful, occasionally sentimental - is the shape of the person behind Exla.
06 / The throughlineCut the waste, keep the signal
Step back and the career reads as one continuous idea told three ways. At Amazon, it was shrinking models so search could stay fast. At Exla's start, it was shrinking those models further so they'd fit on a chip in your hand. Now, with Libra, it's shrinking the distance between a failure and its fix - collapsing hours of frantic human debugging into an automated pull request that arrives before anyone has finished their coffee.
The aspiration underneath all of it is simple and a little audacious: production software that heals itself. Systems that notice their own regressions, trace them to the responsible change, and repair the damage without waking anyone. If Das is right, the dreaded 3 AM page doesn't get better. It disappears.
Exla is still small - two people, a seed check, and a very large ambition. But small teams have shipped bigger surprises, and this one has an unusual asset: two founders who have spent their careers thinking about optimization and about what machines do in the dark. Whether Libra becomes the on-call engineer that never sleeps is an open question. That Viraat Das will keep building toward it is not.