In 2018, Antoni Rosinol and Sebastien Mannai entered an engineering competition with a curious proposition: watch the top of an oil well closely enough, and you might learn what its pump was doing underground. The equipment was already there. So were the cameras. The trick was to make their pictures useful.
The two MIT graduate students had 24 hours to develop a software concept for the Siemens FutureMakers Challenge. Their approach combined machine vision with signal processing, reconstructing the well’s surface motion in three dimensions and using it to infer movement below. They won first prize. Siemens committed $140,000 to further research.
It is a useful place to begin Rosinol’s story because it contains a problem he would keep encountering. Information exists. A machine can collect it. Somewhere between the two lies the difficult business of understanding enough to do something worthwhile. An oil well makes that problem unusually literal: the action you care about is out of sight.
The camera and the oil well
Rosinol would eventually become co-founder and CEO of StackAI, a company connecting artificial intelligence with enterprise data and software. That description sounds a long way from an oil pump. Follow the engineering, however, and the distance shrinks. Both projects ask what useful action requires beyond a stream of information.
The 2018 idea also had an appealing economy. Instead of asking a business to replace everything it owned, the students proposed extracting more value from existing CCTV footage. There is a particular satisfaction in making an ordinary camera earn its keep. The camera has been watching all this time; someone finally gives it a better assignment.
At MIT, Rosinol worked with Luca Carlone in the SPARK laboratory. His doctoral subject was spatial perception, culminating in a thesis on real-time dense metric-semantic SLAM. The degree, awarded in February 2023, was in Aeronautics and Astronautics. The vocabulary is formidable. The underlying ambition is easy to recognize: help a robot understand where it is and what surrounds it.
SLAM means simultaneous localization and mapping. A robot needs to estimate its position while constructing a map. Add semantic information and the map can describe what things are, as well as where they sit. Anyone who has searched a crowded room for a particular chair can appreciate the distinction between having coordinates and having useful directions.
A robot needs a map. So does a company.
Rosinol co-authored Kimera, an open-source library for robot localization and mapping. Its ingredients included camera images and inertial measurements. Its output brought geometry together with labels: a reconstructed environment whose contents could be identified. The software’s name borrowed from a mythical creature assembled from different animals, an apt choice for a system combining different kinds of understanding.
His 2020 work on 3D Dynamic Scene Graphs, with Arjun Gupta, Marcus Abate, Jingnan Shi and Carlone, took the representation further. Objects, people, rooms and structures became entities with relationships. A map could support questions and planning. The research included moving people, because the world inconveniently declines to hold still for the benefit of its machines.
There is a family resemblance between that work and the software Rosinol later helped build. A company also contains entities and relationships: a document belongs to a project, a customer appears in a database, a request belongs to a team. That comparison is an interpretation of his career, rather than a claim that office software and robotic navigation are the same technology.
It helps explain why the office is an interesting destination for a robotics researcher. An intelligent system needs usable context. In physical space, that context may be a room and its contents. At work, it may be a contract, a permission setting and the record of who agreed to do what. The scenery changes. Connecting knowledge with action remains difficult.
Five years before the pitch
Rosinol grew up in Tarragona. His education took him through Switzerland, Singapore and the United States: mechanical engineering at EPFL, an undergraduate exchange at Nanyang Technological University, and a master’s in Robotics, Systems and Control at ETH Zurich. These were distinct stages of training, preceding the doctoral work that later became central to his public biography.
His early career included aerial robotics software at GoPro in Zurich and research at the University of Zurich’s Robotics and Perception Group. He also co-founded Velohub, serving as CTO and working on Blinkers, smart bicycle lights. Before enterprise AI, there was the rather concrete assignment of making a bicycle communicate more clearly with the traffic around it.
In 2021, he worked at NASA’s Jet Propulsion Laboratory on the CoSTAR team participating in the DARPA Subterranean Challenge, with a focus on perception and state estimation. Cameras, movement, uncertain surroundings: the technical concerns were consistent even as the settings changed. His experience was accumulating around machines expected to function beyond a neatly prepared demonstration.
MIT supplied a collaborator as well as a research environment. Rosinol and Bernardo Aceituno began exploring business ideas together. Their earlier entrepreneurial work included warehouse automation; a 2021 MIT delta v presentation listed them on a team developing a modular system intended to fit small, changing warehouses. Entrepreneurship arrived through experiments, with more than one product direction along the way.
When Rosinol later spoke to fellow scholarship recipients about starting a company, he said he and Bernardo had worked together for five years before launching it. His advice placed the founding partner ahead of the idea. It is an unusually personal item in a discussion that could easily have dissolved into market sizes and fundraising tactics.
A partnership has to survive ordinary work before it can carry extraordinary expectations. Five years gives two people time to discover how the other builds, responds and disagrees. Rosinol’s account makes the collaboration part of the company’s substance. The names on the pitch were attached to a relationship that had already been tested by doing things together.
“What’s the worst thing that could happen?”
Antoni Rosinol, discussing entrepreneurship in 2024
That question captures the practical tone of his advice. He encouraged aspiring founders to examine their fear, find complementary people and begin with something a customer might pay for. It brings entrepreneurship down to a manageable scale. The first task can be smaller than the future company, even when the ambition eventually becomes considerable.
The customers you choose to leave
StackAI’s company history dates its founding to 2022. It entered Y Combinator’s Winter 2023 batch and launched publicly in April 2023. Rosinol and Aceituno were building as language models brought a new set of possibilities into view. The gap they pursued was integration: connecting those models with the proprietary data and systems inside companies.
The product used a visual workflow builder. Teams could arrange components and connect business information with models, instead of assembling every application from scratch. That interface matters. A business process often crosses several tools, and the people who understand the process may have little appetite for becoming software engineers merely to improve it.
Early adoption came from a broad audience. Then the company chose a narrower commercial focus. In his Samaipata interview, Rosinol described the move toward enterprise customers as a decision that meant leaving behind consumer and smaller customers representing 70 percent of the earlier base. He reported an elevenfold revenue increase after that shift.
Those figures describe his account of the pivot. They also make the decision more interesting than a routine claim of growth. A signup total can flatter a company while concealing competing demands. Serving a large organization changes the sale, the product and the support required. Concentration asks a founder to surrender some visible activity before learning what the new focus will yield.
By May 2025, StackAI had announced a $16 million Series A. At that point, the company reported more than 90,000 individual signups and over 100,000 agents created. Its announced capabilities included connecting business data with models and deploying workflows for tasks such as preparing investment memos, responding to requests for proposals and using internal documentation to train staff.
Earlier customer base left behind → reported revenue increase after the shift.
Different measures. They illustrate the trade-off, not a causal calculation.The office gets a new map
By March 2026, the founders were articulating a more explicit philosophy. StackAI’s manifesto imagined enterprise processes run by agents, with those agents governed by the teams responsible for them. Its emphasis was ownership, auditability and control. The ambition was broad; the proposed conditions were deliberately specific.
That emphasis follows naturally from the customer choice. A working process needs someone responsible for its consequences. A model’s fluency cannot assign that responsibility. A team needs to understand the workflow, decide what access it receives and retain the ability to intervene. The administrative details may lack theatrical appeal, but they determine whether a promising system can remain useful.
On May 28, 2026, Asana announced that it had completed the acquisition of StackAI. Both co-founders joined Asana. In its earnings call, Asana described approximately $75 million in upfront cash consideration and an additional equity-based earn-out opportunity. That is a transaction milestone for the company Rosinol helped build, with another phase of work attached.
The proposed combination pairs StackAI’s ability to execute workflows across systems with Asana’s project context, ownership and coordination. Rosinol described joining Asana as a way to scale the offering. The attraction is understandable: an agent that can take an action still needs that action to make sense within a team’s work.
The oil-well competition now looks like a modest early chapter in a larger engineering career. Its lesson was concrete enough to survive the change of setting. Useful information can be close at hand and still difficult to act on. Rosinol has moved from cameras and spatial maps to enterprise workflows, carrying that problem into offices whose software has accumulated over years.
There is room for a little sympathy for the machines. First we ask them to understand the world. Then we introduce them to the company filing system. Rosinol’s work concerns the connections that make the assignment possible: between a picture and a pump, a map and a route, a model and the place where a job actually gets done.