From TileDB to Tile.ai
Nine years ago I started a company around a problem most people preferred to route around: the world’s most valuable data didn’t fit the tools built to manage it. Today that company is becoming Tile.ai, I’m stepping into founder mode to focus on product and vision, and we’re welcoming George Llado as CEO. This is the story of how we got here — what we solved with TileDB, what AI changed, and why I decided the company had to change with it.
The problem we set out to solve
I came to this from research — at MIT and Intel Labs — and I kept hitting the same wall from different directions. A genome, a microscope slide, a LiDAR scan, a stream of sensor readings, a machine-learning feature set: none of it fit neatly into a relational table, and each field had grown its own silo of specialized, mutually incompatible tools. Every domain was re-solving storage, access, and scale from scratch, badly, and then copying data between systems just to get work done.
The bet behind TileDB was that there is a single, honest abstraction underneath all of it: the multi-dimensional array, dense and sparse. A table is an array. An image is an array. A genome, a point cloud, a time series — all arrays. If you build one engine that represents them all natively, with the performance and scale each demands, you stop needing a different silo for every data type, and you stop copying data around to make it usable.
That’s what we built, and I’m proud of what it solved:
- One engine for many modalities, instead of a fragmented tool per data type.
- Performance and scale that specialists in genomics, imaging, and geospatial actually trusted — cloud-native, with data living on object storage and compute running anywhere.
- An open format, published and unencumbered, rather than a proprietary black box.
- Governance, versioning, and sharing at the data layer, not bolted on after the fact.
TileDB became the infrastructure for some of the most demanding data work there is — population genomics, single-cell, imaging, geospatial, ML — in exactly the regulated, high-stakes environments where getting data wrong is not an option. It earned recognition from analysts and, more importantly, the trust of teams doing serious science. The open-source engine will keep going, under the TileDB name, exactly as it is.
The principles that turned out to matter
Building that engine for years taught me a few things I now hold as close to first principles:
Migration is a tax most people won’t pay — and shouldn’t have to
I learned this one in reverse, from our own architecture: however good the destination, asking people to convert their data and give up the tools built on their native formats is a cost many simply can’t justify. Data should be met where it already lives, in whatever format it is already in.
A single interface matters more than a single format
The moment you can reach wildly different data through one coherent interface, whole categories of glue code and integration pain disappear — and, I eventually realized, you can have that single interface without forcing everything into one store.
Governance belongs at the data layer
Access, lineage, and control that are properties of the data itself survive; the kind bolted on top by each application leaks.
Some of these I designed for from day one. The most important one I learned the hard way. All of them turned out to be exactly what the next era would demand.
Then AI changed the question
For most of TileDB’s life, customers came to us with a performance question: make my complex data faster and bigger than anything else can. That never stopped mattering — it never will. But over the last two years, a second question arrived on top of it, and it quickly became the one that kept people up at night: make my data usable by AI — safely, where it already lives, without losing control of who can see what.
The models had stopped being the bottleneck. The data had become it. Enterprises discovered that their most valuable, proprietary data was scattered across files, databases, instruments, cloud services, and code — inconsistently governed and barely reachable by an AI agent in any trustworthy way. Pilots stalled not because the models weren’t smart enough, but because the data underneath them couldn’t be reached, used, or trusted.
And the old fragmentation problem came back, worse. It was no longer enough to store and query complex data well. Now every one of those sources had to be made usable by an autonomous agent: with the real tools to work the data in its native form, the access rules that keep it safe, and — critically — the ability for the agent to act as the actual person asking, not some shared, all-powerful service account. Generic connectors hand an agent raw bytes. Catalogs only describe what exists. What was missing was a layer that makes any source genuinely usable by AI, in place, under real identity and real governance — and that can actually compute on the data, not just point at it.
When I stepped back and looked at it, the answer was almost uncomfortable: the principles we’d spent nine years hardening — one coherent model across modalities, governance at the data layer, real performance at scale — were precisely what this new era needed. A storage engine, however good, wasn’t the whole answer. The product had to grow beyond TileDB arrays into a layer that spans every system an enterprise runs — meeting data in whatever source and format it already lives in, turning each source data asset into a governed, AI-ready Tile, and letting agents reason across many of them as one governed whole rather than a dozen disconnected pieces.
That layer is Tile.ai — the Enterprise AI Data Substrate. It carries the conviction I started with — that an organization’s hardest data should be fully usable — but with the lesson TileDB taught me built in from the start: meet the data wherever it lives, in the format it is already in, and never make anyone migrate to get value.
Why I’m stepping into founder mode
Recognizing the opportunity is the easy part. The hard part was admitting what it required: not a feature, not a new logo, but a company reshaped around a new mission. That is a full-time product and vision problem, and it deserves someone giving it everything.
So after nine years as CEO, I made a deliberate choice to step into founder mode again — hands-on alongside my fantastic team, focusing on the product and the vision, the work I love most and the work this moment demands. It’s the part of the job I’m best at and most needed for right now.
That decision was only possible because we found the right person to run the company. I’ve worked with George Llado for the past two years, and I’ve watched him operate. As a longtime CIO in life sciences, he lived this exact problem from the other side — responsible for making data useful while keeping it safe in one of the most regulated industries there is. He knows the buyers, the constraints, and the stakes better than almost anyone. George embracing our vision and accepting the CEO role to drive our commercial growth in those industries is one of the highlights of my career.
I’ll be honest about where we are: this is the beginning, not the finish line. Tile.ai is in private preview, and we’re building it shoulder-to-shoulder with a small set of design partners in regulated industries. That’s on purpose. The last thing this problem needs is another platform designed in a vacuum and thrown over the wall.
Why I’m more excited than I’ve ever been
The through-line of my entire career has been a single stubborn belief: that an organization’s hardest, most valuable data should be usable — fully and safely — no matter how complex it is or where it lives. For nine years, that meant making it fast and scalable, even when that meant bringing the data to us. The lesson I carry into Tile.ai is that it shouldn’t have to move at all — now the goal is making that data usable by AI, right where it already is, on the enterprise’s terms.
Same belief. Much bigger stakes. And, for the first time, the whole industry is finally asking the question we’ve spent a decade preparing to answer.
And this conviction isn’t mine alone anymore. Amgen — a leader in one of the most data-intensive, regulated industries there is — chose to back this transition, and Two Bear Capital, who’ve known us for years, doubled down. When a strategic partner who lives this problem every day, and investors who’ve watched us up close, both decide the pivot is right, it validates more than the roadmap — it validates the hypothesis I reshaped the whole company around. If you’re wrestling with this — proprietary data everywhere, AI ambitions stalled on governance and access — I’d genuinely like to hear from you. We’re building Tile.ai with people who live this problem, and there’s room at the table.