Jay Ram Wants to Change the Rules of the Game

Jay Ram had already found an audience for artificial intelligence. It wanted to make prank calls. Before becoming the founder and CEO of HUD, he built an AI prank-calling app that, by his account, briefly reached number one on the App Store and generated roughly half a million calls. There are founders who would happily spend the next several years defending a position like that. Ram went looking for a different problem.

The episode is a useful entrance to his story because it complicates the usual arithmetic of startup success. A chart can tell you that people want something. It cannot tell you whether you want to spend your working life making it for them. Ram’s next move took him toward researchers, computer-use agents and reinforcement learning. The audience became smaller, more technical and rather less interested in a good telephone joke.

Today his company supplies infrastructure for building and evaluating the environments in which AI agents attempt tasks. HUD also operates a marketplace where builders can sell training data to AI labs. Underneath the technical vocabulary sits a recognisable human question: how do you arrange the conditions so that someone, or something, gets better? Ram has made that question his business.

A student with several tabs open

Before HUD, there was Columbia University. Ram’s public education record lists 2020 to 2023; his 2023 8VC fellowship biography describes a senior majoring in computer science and physics. The combination puts code beside an older habit of asking how a system behaves. It is a useful pairing for someone whose eventual product would involve both running experiments and deciding what the experiments should mean.

His interests were already spread across several kinds of systems. He had interned in quantitative finance, applying statistics and machine learning to financial modelling. At Standard Metrics, he worked on its POST API, used by venture capital firms including 8VC. He also researched large language model interpretability and ran a fellowship concerned with economic progress. These were different projects with different audiences, rather than one tidy undergraduate specialism.

Outside those projects, the same biography lists reading, hackathons and mixed martial arts. None proves a theory about his character. Together they offer a more useful picture than the frictionless founder biography: a student who could move between a book, a code project and a setting where an opponent had immediate feedback to offer. Algorithms are fortunate to receive their corrections in numerical form.

Quantitative finance supplied an established career path. Ram left it to build products. The revealing detail in his account is the dissatisfaction with merely talking about what he might make. A weekend conversation can accommodate an extraordinary number of imaginary companies. A shipped product has customers, complaints and the awkward advantage of existing. His prank-calling app put that difference into practice.

The customers changed the question

HUD brought Ram together with Lorenss Martinsons and Parth Patel, its fellow founders. The company joined Y Combinator’s Winter 2025 batch. Its founding team describes the original frustration as one encountered while building computer-use agents at Hume: the agents were advancing, but the tools for training and evaluating them were lagging. The missing infrastructure became the work.

An agent that operates software needs more than a plausible answer in a chat window. It needs to complete a task through actions. That creates awkward questions. Did it reach the right result? Did it accidentally take advantage of a broken test? Can the same attempt be run again under comparable conditions? A convincing demonstration is pleasant. A repeatable experiment is more useful to the person trying to improve the system.

HUD’s early work concentrated on evaluations. Its January 2025 Autonomy release described more than 100 tasks spanning nine domains, including documents, browsing, applications, collaboration and games. The team reported that expert humans completed more than 98 percent of the tasks, while the best AI agents completed less than 25 percent. Those are historical results from that evaluation, rather than a verdict on what today’s models can do.

The gap gave the company a concrete place to begin. But Ram’s account of the customer conversations contains the more consequential development: requests for evaluations were also requests for environments. Researchers needed somewhere to run an agent, give its actions a reward and train on what happened. The customer’s vocabulary was pointing at a larger job than the original label suggested.

A score needs somewhere to go

An evaluation records how a system performs. A training environment gives the system conditions in which performance can change. In reinforcement learning, the reward tells the training process which outcomes to favour. That makes the definition of success part of the engineering. If the reward accepts the wrong result, getting better at collecting it can make the agent less useful.

Consider a deliberately simple example: an agent is asked to export a report. A check that merely notices a file may approve an empty export. A more useful check inspects whether the required rows arrived. This is an illustration of the problem HUD’s current product addresses: inspecting traces, reviewing tasks and finding mistakes in graders. An agent can pass a test while the person who requested the work still has nothing to use.

For Ram, this is also a product-design problem. In his interview, he describes HUD using the phrase “Next.js of RL environments,” borrowing the name of a web framework to explain the kind of development experience he wants. The appeal of that comparison is practical: give builders an organised way to work, with useful decisions already made, while allowing them to connect to different training backends.

The less glamorous part is speed. An experiment that returns useful information quickly gives its builder another chance to revise the task, change the reward or inspect a failure. Long waits make each wrong assumption expensive. Ram’s product argument depends on shortening that loop. The object being sold is infrastructure, but a good part of its value is the next experiment a researcher has time to run.

“Next.js of RL environments”

Jay Ram’s description of HUD

The board has 32 seats

Then there is Catan. In July 2026, HUD and Fireworks AI organised a tournament in San Francisco with a 32-player bracket and roughly $3,500 in prizes. The event description offered an unusually relevant piece of personal colour: Ram had been rated among the top 20 players in the United States and was due to compete. The listing did not specify the date or system behind that ranking.

The tournament grew out of a project at the previous month’s HUD hackathon. A player described by the organisers as number one on Colonist.io had used HUD and Fireworks AI to train a model to play Catan. A company working on reinforcement learning had found a reason to put people around a game board. For once, the networking event came with actual trading.

It would be easy to make the game explain the entrepreneur. Too easy. A board-game ranking does not establish how a person runs a company. The better detail is the overlap: Ram’s working world and recreational world both contain rules, repeated attempts, competing strategies and outcomes that can be examined. At the tournament, the CEO was listed as a participant, with a seat in the bracket rather than a ceremonial exemption.

Making expertise saleable

By June 2026, HUD’s ambitions included a market as well as a toolkit. The company’s funding announcement reported more than 50 businesses using the platform and described a $16 million Series A led by Standard Capital. It listed Y Combinator, Exceptional Capital, Liquid 2 Ventures and Twenty Two Ventures among the participants. Those are company disclosures, giving a dated picture of the business rather than an independent measurement of its prospects.

The marketplace proposition follows from a division of knowledge. AI labs know how to train models. A person inside a particular profession may know what a difficult, realistic task looks like, what a convincing result requires and which apparently correct shortcuts are useless. HUD wants those people to be able to package that expertise into environments and data that labs can purchase.

That changes the kind of customer Ram is pursuing. The prank-calling app offered an immediate experience to a consumer. An environment builder has to produce something another technical organisation can inspect and trust. The transaction depends on a chain of judgments: a task worth learning, an environment that represents it, and a grader that recognises meaningful success. Each link can break. Each offers work for the company.

Ram’s longer view includes more specialised models, adapted to the domains where businesses actually operate. In that view, teams with an established product may choose to train for better accuracy or lower cost. This remains an argument about where the market is going. Its appeal is concrete, though: the people who understand the work should have a way to influence what the model learns.

The movement through his career is striking without needing to be smoothed into destiny. Finance, consumer entertainment and agent infrastructure reward different things. Ram has tried each and kept changing the problem he wants to sit with. His current wager is that teaching an agent requires an environment worth learning from. The rules deserve as much attention as the player. He has chosen to build them.