The bubble-tea machine belongs in the story. So does Jankbot, a homemade robot/go-kart with a name that wisely declines to promise elegance. Before Yash Patil was selling companies on the merits of owning their artificial intelligence, he was making things with unusually clear job descriptions. Dispense a drink. Fetch information by text message. Help students find each other. His portfolio has the charm of a workbench on which nobody has yet imposed a policy about tidying up.
Today, Patil is co-founder and CEO of Applied Compute, working on custom models and the infrastructure to train and run them. The setting has changed considerably. The underlying question remains wonderfully practical: what, exactly, should this thing do? In an industry fond of asking how much a machine knows, Patil has built a company around the work a particular organization needs it to finish.
A phone with no data plan
Consider GimmeSMS, an early project designed for people without cellular data plans. A user could text a number and receive information from the web, including directions, locations, and weather. The service used Python, Twilio, and Google Maps APIs. It took something already abundant online and made it available through a channel a person could actually use. A browser was an assumption; a text message was an alternative.
Another project, Ecuisina, let neighbors barter home-cooked food. Its features included listings, trade requests, scheduling, and searches based on distance and preferences. That is a substantial amount of machinery to put behind the friendly proposition that somebody nearby might make excellent dinner. Both projects took an everyday inconvenience seriously enough to build a system around it.
Patil grew up in Austin. He attended the Liberal Arts and Science Academy before studying computer science at Stanford. His early technical work also included modeling the Texas power grid at the University of Texas at Austin's Webber Energy Group. The research involved simulating generation and transmission and organizing the outputs of different model runs. His interests could accommodate both an electrical grid and a neighborhood food swap.
The work gained organizational weight at Helping Hands, which he co-founded. He built the first version of its user platform and helped expand the engineering team to more than 15. The system connected food-delivery requests with volunteers and logistics providers. Helping Hands delivered more than two million meals across all 50 states. Software here had a destination beyond the screen.
Meals delivered through Helping Hands, the platform Patil co-founded.
A platform total, shared across the organization.Stanford, with extra moving parts
At Stanford, he co-directed the 2023 edition of TreeHacks. His campus projects addressed the small social negotiations that fill student life: finding classmates, comparing dining halls, discovering events, and arranging conversations with people who shared an interest. There is something pleasing about a computer-science portfolio that leaves room for the question of where lunch might be tolerable.
He also co-founded Armadillo Labs, applying machine learning to online storefronts. That venture connected a technical discipline with a commercial problem well before Applied Compute. Meanwhile, physical projects continued to appear beside the software. The bubble-tea dispenser and Jankbot make the list harder to mistake for a résumé assembled solely to satisfy an interview panel.
That collection should resist the temptation to become a prophecy. A homemade machine does not secretly contain the business plan for an AI cloud. What the projects do establish is range: Patil had experience writing software, coordinating people, and building things whose usefulness could be tested directly. A delivery either arrived or it did not. A classmate either joined the conversation or remained somewhere else.
A dinner, then an application
His route to OpenAI included a dinner he hosted for Stanford engineering students, with Sam Altman among the guests. Patil later described being nervous and barely speaking to him. After ChatGPT appeared, he emailed Altman about working on it. Altman encouraged him to apply to the residency program. The useful part of that encounter happened after the dinner plates had gone away.
Patil joined OpenAI at 20 and took leave from Stanford. He worked on post-training infrastructure and the effort that became Codex. Rhythm Garg and Linden Li, friends from Stanford, joined OpenAI too. Patil referred Garg; Garg referred Li. Their later founding team grew from relationships that had already crossed a campus and a research lab.
Their work covered complementary pieces of the problem. Patil worked on the coding-agent effort, Garg contributed to o1, and Li worked on machine-learning systems for reinforcement learning. When they founded Applied Compute in 2025, the three brought experience with agents, reasoning, and the machinery required to train them. Patil became CEO, Garg CTO, and Li chief architect.
The intelligence inside a company
Applied Compute's public introduction in October 2025 named Cognition, DoorDash, and Mercor as early customers. It announced $80 million in backing. By April 2026, the company disclosed another $80 million round, led by Kleiner Perkins, bringing total funding to $160 million at a $1.3 billion post-money valuation. Those figures financed a proposition with a deliberately narrow unit of ambition: AI built for the work of one company.
Patil calls this “Specific Intelligence.” The argument begins with a familiar organizational fact. Two companies can hire equally qualified people and still operate differently. They have different tools, standards, customers, and accumulated judgment. A general model may possess broad knowledge while remaining unfamiliar with the decisions that make one business function well.
Applied Compute works with those businesses to train models around their own data and definitions of success. Its engineers work alongside customer teams. The commercial premise is that valuable expertise already exists inside an organization, distributed across people, examples, and corrections. Turning it into something a model can learn requires close attention to how work is actually judged.
“No one is data ready,” Patil said in a June 2026 conversation. It is a useful remark from someone whose company depends on helping organizations use their data. A folder full of records and a training process are different achievements. Before the model can improve, the people involved must decide what deserves a good score.
“No one is data ready.”
Yash Patil, June 2026
The menu is the examination
DoorDash supplies an example refreshingly resistant to abstraction. Its merchant-onboarding system turns photographed menus, PDFs, and text into structured listings. Real menus are messy. Layouts vary, and experienced reviewers can disagree about the right interpretation. The task requires a definition of quality close enough to the actual business to be useful.
Applied Compute and DoorDash worked on an automated grader calibrated to expert judgments, then trained against that standard. In February 2026, the companies reported a 30% relative reduction in critical menu errors. The qualification matters: this was an improvement against a baseline for a particular workflow. It gives Patil's thesis a concrete test, with fewer restaurant listings requiring correction.
Reinforcement learning supplies a method for making those improvements. A model attempts tasks; a grader evaluates the attempts; training adjusts the model toward behavior that receives better rewards. The arrangement makes the grader consequential. Reward the wrong behavior consistently and the training process can become an efficient way to manufacture the wrong result.
- AttemptRun the agent in a task environment.
- EvaluateScore the work against a defined standard.
- TrainUpdate the model using that feedback.
- Use + repeatBring production experience into another run.
Patil's attention extends to the less photogenic requirements underneath. Thousands of attempts need environments in which to run. Grading must keep up. Expensive GPUs should spend their time doing useful computation. In explaining the company's infrastructure, he talks about parallel execution, reliability, and matching training environments to the systems agents will encounter at work.
A workshop other teams can enter
On August 25, 2026, Applied Compute introduced AC2, its Agent Cloud. The platform made the company's internal training and inference tools available to other AI teams. Customers could bring their own agent harnesses, inspect training attempts, deploy models, and use production feedback in subsequent experiments. The business was extending from embedded work into a platform other researchers could operate.
The distinction is consequential for Patil's job. Helping one customer improve one model demands expertise. Giving many customer teams a repeatable way to do it demands a usable system. AC2 brings the company's claim about model ownership into the practical territory of experiments, deployment settings, and the next training run.
October brought another expression of that interest in practical economics: Shift-1, a router that chooses models for individual requests inside a coding session. Applied Compute reported that internal use reduced Codex spending by 55% against its all-GPT-6-Astra comparison, with no discernible quality loss by its evaluation. It is the company's own result, measured on its own work, rather than a promise about every customer's bill.
Routing also makes Patil's position more interesting than a simple allegiance to one type of model. A task can need different capabilities at different moments. Choosing where to spend computation becomes part of the product. The question returns to fit: which model, for which request, at what cost, under whose standard of success?
That is a long way from a bubble-tea dispenser, although both projects benefit from a clear specification. Patil's current ambition is to give companies the means to train, run, and improve intelligence they control. The next test will arrive in the work itself: an agent's attempt, an expert's correction, a model that learns something useful. The machine still has a job to do.