The business world has developed a peculiar appetite for verbs ending in “maxxing.” They sound muscular, urgent and conveniently exempt from questions. In artificial intelligence, the latest object to be maxxed has often been tokens: more prompts, more model calls, more computational confetti. Arsalan Tavakoli-Shiraji, a co-founder of Databricks and the company’s senior vice president of Field Engineering, would like to audit the party.
His alternative is “value maxxing,” a phrase light enough for a conference stage and stern enough for a finance meeting. Which model is right for this job? What will it cost? Who can use it? Can its decisions be governed? Above all, did anything useful happen? Tavakoli-Shiraji has arrived at these questions by an unusual route: computer science and economics at the University of Virginia, a doctorate in distributed systems at UC Berkeley, strategy work at McKinsey, and then the founding of a company devoted to making unruly data usable.
It is tempting to file him under “technical founder” and leave the folder shut. That misses the interesting part. His career has been a long commute between the machine room and the meeting room. Field engineering is where those rooms collide. The job is not merely to explain software. It is to sit with a customer, discover where the grand architecture meets a cranky permission system, and make the promised outcome occur.
A doctorate about getting from here to there
In 2009, Tavakoli-Shiraji completed a dissertation with a title only a network engineer could love: “Exploring a Centralized/Distributed Hybrid Routing Protocol for Low Power Wireless Networks and Large Scale Datacenters.” Scott Shenker advised the work. Its two settings sit at opposite ends of the computing landscape, tiny constrained devices and enormous data centers, yet the connecting problem was routing. How should a system move information reliably when resources and conditions differ?
The title now reads like an accidental preview. Enterprise technology is also an exercise in routing: research into products, products into organizations, organizations toward results. The elegant answer drawn on a whiteboard must survive old databases, security rules, budgets and human impatience. A distributed-systems education offers a useful suspicion of perfect conditions.
After Berkeley, he joined McKinsey and rose to Associate Principal. He advised enterprises, technology vendors and public institutions on cloud computing, next-generation IT and corporate strategy. The shift from network protocols to consulting decks might look abrupt. In retrospect, it supplied the other half of his working vocabulary. He could talk about the system and about the institution expected to buy, govern and change around it.
“I don't think it makes sense to focus on technology for technology's sake.”Arsalan Tavakoli-Shiraji, 2025
Seven founders and one unfashionable bet
Databricks began in 2013 with seven co-founders connected to the Berkeley research community: Ali Ghodsi, Ion Stoica, Matei Zaharia, Patrick Wendell, Reynold Xin, Andy Konwinski and Tavakoli-Shiraji. Apache Spark, the open-source engine born at Berkeley, had made large-scale data processing faster and more approachable. The company would take that research-rooted work into the market.
Customers did not greet every premise with rose petals. Tavakoli-Shiraji has recalled the pressure to put the product on premises. People told the team that the money was there, not in the cloud, and that serious organizations would never place real data in somebody else’s infrastructure. It was practical advice with a short shelf life. The founders kept the cloud focus.
“Sticking to your guns despite opposition, if you feel strongly, is a really big deal,” he later said in a summit conversation. Conviction in this telling is not mystical. It is a decision maintained through an awkward interval, after the consensus says no and before the evidence says yes.
His answer became Field Engineering, the technical organization that works alongside customers. In a 2025 recruiting note, he put its size at more than 3,000 people, roughly a third of Databricks at the time. More revealing than the number was the job description. He wanted a “hands-on-keyboard builder mentality,” not a conventional enablement apparatus that broadcast training from a distant stage.
His preferred methods included hackathons, in-context learning, personalized AI tutoring and the codification of technical practices without erasing local differences. This is the scaling problem in miniature. A small group can learn by sitting beside one another. A global group needs systems. The trick is to build those systems without turning builders into spectators.
The glamour of a solved invoice
This operator’s view explains Tavakoli-Shiraji’s fondness for “boring AI.” The phrase is a compliment. Fraud detection is boring until it stops a fraud. Predictive maintenance is boring until a machine stays online. A faster onboarding process lacks the theatrical possibilities of a talking robot, but the saved hours appear in a ledger.
He makes a careful distinction between the technology used and the outcome achieved. Sometimes the appropriate answer is an AI agent. Sometimes it is traditional machine-learning automation. No customer should brag about how many agents it deployed, any more than a restaurant should advertise the number of pans it owns. The useful report is that a process used to cost this much and take this long, and now costs less or moves faster.
In Korea, he reduced the foundation to two words: data and governance. Without them, he said, a company “doesn’t have a prayer of getting AI right.” The local details mattered. Some Korean businesses still kept significant data on premises, partly because of national-security considerations. Large conglomerates also had to impose consistency across global operations. The AI model was only one guest at a very crowded table.
The same habit appeared in a 2026 discussion with Visa technology president Rajat Taneja. The difficult subjects were permissions, auditability, production testing and process redesign. Agents cannot simply inherit human-level access because they appear human-like in a chat window. The software stack changes when an agent replaces an interface, and governance has to change with it.
Their advice to founders ended on a wonderfully practical note: throw out the PowerPoint and show the product. This may be the purest expression of field engineering. Slides describe a possible future. A demonstration creates a small, testable piece of it.
“At the end of the day, it comes down to two things: data and governance.”Arsalan Tavakoli-Shiraji, Seoul, 2025
The field is the point
Tavakoli-Shiraji is an Iran-born founder in a company assembled from an unusually international Berkeley group. His public profile is quieter than those of some co-founders. He does not cultivate a grand personal doctrine with a trademarked diagram. The ideas recur instead through customer forums, recruiting posts and interviews: remain technical, start with the problem, preserve choice, govern the system, measure what happened.
That consistency has become more relevant as AI moved from laboratory spectacle into corporate infrastructure. In 2024, he discussed the next AI frontier in Singapore. In 2025, he spoke in Seoul about cloud adoption and consistent governance. In 2026, India’s digital businesses and global capability centers were part of the conversation. The geography changes. The bottlenecks rhyme.
The Seoul discussion supplied a crisp picture of where he believes the value will collect. Infrastructure forms the base of an AI pyramid. Model providers occupy the layer above. Applications sit at the top, still smaller today, but positioned to deliver most of the eventual value. It is a field engineer’s pyramid: impressive machinery at the bottom, a result that somebody can recognize at the top. The diagram also explains why he resists beginning with the name of a model. Buyers live at the application end. They care whether the warehouse predicts demand, the factory improves yield or the bank catches fraud. The rest is necessary plumbing, and plumbing is judged most kindly when it works.
His newest phrase, the shift from token maxxing to value maxxing, catches a maturing market. Companies now ask how to control spending, govern AI use and route each task to an appropriate model. The largest model is not automatically the wisest choice, just as the loudest consultant is not automatically invited back.
There is also a modest philosophy of progress here. New technology rarely creates value by being sprinkled over an old process. Work has to be reconsidered. Permissions have to be redrawn. People have to be taught and trusted. The less photogenic work around the model often decides whether the model matters.
That brings Tavakoli-Shiraji back to the place where his career makes the most sense: the field. A field engineer stands close enough to the customer to hear the complaint, and close enough to the product to do something about it. After the keynote, the funding announcement and the astonishing demo, somebody still has to put hands on a keyboard. For him, that is not the anticlimax. It is the entire point.