Profile / Data Infrastructure
Manu
Bansal.
He spent his last startup debugging bad data. So he built a company to make data tell the truth.
Manu Bansal - two companies, one grudge against bad data.
He spent his last startup debugging bad data. So he built a company to make data tell the truth.
Most engineers building artificial intelligence want to talk about the model. Manu Bansal wants to talk about the moment the model gets fed a number that is quietly, catastrophically wrong - and nobody notices until the dashboard is on fire. That moment is the reason his company exists.
Bansal is the co-founder and CEO of Lightup, a data-quality and observability platform built for the messy, sprawling data stacks that large enterprises actually run. The pitch is unglamorous on purpose. Lightup watches the pipes. It runs hundreds of thousands of automated checks a day across petabytes of customer data, and it raises its hand when something drifts, breaks, or goes silently stale - before a bad row reaches a board deck, a billing run, or a machine-learning model that will faithfully turn garbage into a confident prediction.
He arrived here the hard way. Before Lightup, Bansal co-founded Uhana, a machine-learning analytics company for mobile carriers that grew out of his Stanford doctoral research. Uhana was sharp engineering applied to a brutal problem: predicting and optimizing the behavior of carrier networks in real time, on a firehose of telemetry. It worked well enough that VMware bought the company in 2019. But the building of it left Bansal with an itch he could not stop scratching.
The itch was this. He had set out to build predictive AI for telcos. Instead, he found himself spending most of his days debugging unexpected data problems - missing fields, schema changes, values that looked plausible and were anything but. The intelligence was the easy part. Trusting the inputs was the war. When the dust settled on the acquisition, he did not chase a flashier AI idea. He went straight at the thing that had been eating his time, and made it the mission.
A copilot approach works way better than an autopilot setting.
Trace the line backward and the throughline is consistency, not reinvention. Bansal's engineering education spans three of the most selective programs on the planet: a B.Tech and M.Tech in computer science from IIT Kanpur, a master's in computer science at UCLA, and a PhD in electrical engineering at Stanford. The Stanford years were the hinge. His doctoral work sat in the group of professor Sachin Katti, and when it came time to commercialize that research, Bansal did something most students never do - he co-founded a company with his own advisor.
That company was Uhana, founded in 2016. It took academic work on optimizing networks with machine learning and pointed it at telecom carriers, who sit on some of the largest and fastest-moving datasets in any industry. Uhana built predictive analytics that ran in real time. The technology was good enough that VMware acquired it in 2019 to bolster its carrier and telco portfolio. For a first company born out of a thesis, that is a clean result.
But Bansal kept returning to the same uncomfortable observation. The AI was not the bottleneck. The data was. He has described building tremendous real-time systems for telcos and then spending most of his energy chasing down data issues that had no business being there. It is the kind of insight that is easy to mutter and hard to act on, because the fix is unsexy. Nobody writes breathless headlines about a tool that confirms your numbers are correct. Bansal built it anyway.
Lightup's core argument, in Bansal's framing, is that checking whether data is trustworthy is fundamentally different from checking whether a server is healthy. Infrastructure monitoring asks about CPU and memory - finite, well-understood signals. Data quality asks something closer to a business question: is this metric behaving the way the business expects it to behave? Bansal calls these Data Quality Indicators, and he is blunt that defining them well takes creativity, not just automation. You are modeling the shape of correct, and correct is specific to each company.
That belief shows up in how the product is built. Bansal is wary of the fantasy that you can fully automate trust. "Every enterprise data model is unique and specific to the business," he has said. "There are always nuances. There are always corner cases." The corner cases are where naive automation quietly fails, and where a vendor promising magic loses the customer's confidence the first time it cries wolf.
If there is a single sentence that captures how Bansal thinks about the current AI moment, it is this: "A copilot approach works way better than an autopilot setting." He is not against automation - Lightup runs at a scale that would be impossible without it. He is against removing the human from the loop too early. His view is that automation delivers most when there is a safety net under it, when supervision stays accessible enough that the system can scale without quietly losing the plot. "That's when automation really delivers," he says, "with that safety net."
The second hard-won lesson is about people, not pipes. Bansal describes Lightup's most successful deployments as ones that spread - from a handful of data experts to hundreds of users, some deeply technical, some entirely business-facing. The goal is democratization: making data quality a thing the whole organization can see and shape, rather than a dark art practiced by a few engineers. That is a harder product to build than a clever anomaly detector, because it has to be legible to a marketing analyst and a data engineer at the same time.
Reflecting on five years of building Lightup, Bansal allowed himself a rare moment of accounting. Enterprise customers were running more than 500,000 data-quality checks a day. Coverage spanned over 12 petabytes and 2,500-plus monitored tables, with 500-plus people involved in managing quality. The headline number: a reported 90%-plus reduction in business incidents caused by bad data. He noted that this would have seemed impossible five years earlier - and then, in the same breath, framed all of it as merely "the start."
That posture is worth sitting with. A founder with one acquisition already behind him, running a company backed by Andreessen Horowitz and a roster of data-infrastructure heavyweights, looks at a 90% reduction in bad-data incidents and treats it as a warm-up. It is not false modesty. It is the read of someone who has seen how deep the problem goes, because he lived inside it at his last company before he ever named it.
Bansal's orbit is dense with people who know data infrastructure cold. Lightup's investor list includes Andreessen Horowitz's Martin Casado, himself a former VMware executive and a foundational figure in software-defined networking. Its advisor bench reaches into Databricks, PayPal, Okta, Rakuten, Accenture, and Stanford's faculty. This is not a founder selling a category he just discovered. It is one operating among the people who built the modern data stack and know exactly where its weak joints are.
There is a tidy symmetry to the whole arc. Bansal's first company taught carriers' networks to predict the future. His second makes sure the data feeding those predictions is real. Same instinct, applied one layer down - to the foundation everyone else builds on and nobody wants to think about until it cracks. He is, by temperament, drawn to the part of the system that only gets attention when it fails. He decided to give it attention before it does.
Ask where this goes and Bansal's answer is consistent with everything else about him: trustworthy data at enterprise scale, owned not by a priesthood of specialists but by everyone who depends on it, with automation that knows its limits. In an industry sprinting toward generative AI and autonomous agents, that is a quietly contrarian bet. The more decisions you hand to machines, the more it matters that the data underneath is true. Bansal figured that out the slow way, on his last company's time. Now he is making sure the rest of us do not have to.
A copilot approach works way better than an autopilot setting.
Every enterprise data model is unique and specific to the business. There are always nuances. There are always corner cases.
That's when automation really delivers - with that safety net.
Building data quality checks is more like tracking business KPIs than monitoring CPU or memory.
He co-founded his first company with his own PhD advisor. Most students hand in a thesis. Bansal turned his into a startup, then sold it.
His engineering pedigree spans IIT Kanpur, UCLA, and Stanford - three of the most selective programs on three different campuses.
Lightup exists because of the most tedious part of his last job. He took the chore he hated and made it the mission.
His cap table reads like a data-infrastructure hall of fame - including a16z's Martin Casado, a former VMware exec.
He builds the kind of software nobody notices until it is missing. The highest compliment his product can earn is silence.