Moulik Budhiraja: How He's Building AI That Acts Before a Prompt

The Proletic co-founder and CTO is developing predictive autocomplete for the computer, a system designed to recognize what a user is doing and suggest the next action without forcing them to stop and explain the task first.

The familiar rhythm of using AI begins with a request. A person opens a chatbot, describes what they want, provides enough context for the model to understand the task, and waits for a response. Moulik Budhiraja is working on a different interaction. As co-founder and chief technology officer of Proletic, he leads the technical work behind software that observes computer use, predicts what the person is likely to do next, and places that suggestion directly inside the active workflow. The user can accept the action with Tab rather than interrupting the task to write a prompt.

“The computer already has context about what you are doing,” Budhiraja said. “The question we are working on is whether the system can use that context to offer the next useful action at the moment you need it.”

Building AI around prediction changes the technical problem substantially. Many AI products can begin with a user instruction and send that request to an existing model. Proletic has to determine what is happening before the user asks for help, which requires its own way of collecting computer state, representing that information, training models on it, serving predictions, and carrying out an approved action. Budhiraja’s work spans that entire path, from information gathered by the operating system to the action eventually executed on the computer.

Budhiraja’s role reaches across much of Proletic’s technical stack. He has worked on major parts of the native macOS collection system, the pipeline that converts computer activity into training examples, the model-training and evaluation stack, reinforcement-learning infrastructure, the inference server, and the executor that performs an accepted action. His contributions also include the AXMonitor integration, the company’s internal model-training loop, cleaned-data training, GRPO training and evaluation, and work to stabilize execution.

“I cannot look only at the model and assume the rest of the system will take care of itself,” he said. “The prediction has to come from the right computer state, arrive quickly enough to be useful, and turn into the right action after the user accepts it.”

His work comes as computer-use AI begins moving beyond chat windows. Coding agents can often operate for longer stretches because the task has been defined in advance. Everyday computer work behaves differently. A new message can arrive, a document can change, a user can make a choice, or another application can introduce information that alters what should happen next. Budhiraja sees that constantly changing state as one of the central technical problems for AI systems intended to assist throughout the working day.

For him, usefulness depends partly on eliminating the pause that current AI tools often require. A person should not have to stop what they are doing, open another interface, describe the current situation, and wait for the system to produce a long plan. Proletic instead presents a proposed next action where the person is already working. The interaction is intentionally small because the software is trying to reduce the distance between intention and execution rather than create another destination competing for the user’s attention.

“Every time you have to explain the state of your work to the AI, you are doing translation work,” Budhiraja said. “I am interested in systems that can understand enough of what is already happening to make that interaction much lighter.”

That idea emerged from years of building computer-use systems with Proletic co-founder Jeffrey Zang. Budhiraja studied Software Engineering at the University of Waterloo, and the two later organized ChessHacks, where they built infrastructure that allowed 300 participants to train and run chess models during the event. They had already spent years working on computer-use systems before starting Proletic full-time in December 2025.

Once Proletic began, the work moved quickly from concept into infrastructure. The first code was committed in December 2025, and an early fine-tuned model was trained during the first week. The collection system was rebuilt as a native macOS application that month, and the team shifted to fuller accessibility-state capture in February 2026. By spring, Budhiraja was working across collection, training, evaluation, inference, and action execution as the company developed the system end to end.

The technical challenge extends well beyond predicting a likely click. Computer activity contains information spread across applications, windows, interface elements, changing text, and previous actions. Budhiraja argues that computer-use AI has to understand enough of that structured state to respond at interface speed. That requirement is one reason Proletic built its own collection layer and models rather than relying on screenshots as the basic representation of computer activity.

“The next action only makes sense in context,” Budhiraja said. “If the system has the wrong state, the prediction can be reasonable in theory and still be useless on the actual computer.”

His focus now is moving Proletic’s collection and training stack toward software people can use throughout the day. The technical work includes larger training runs, stronger evaluation against actions users actually accept, reinforcement learning from those approvals, lower inference latency, and more reliable selection of the correct interface element.

Budhiraja sees the larger opportunity in reducing how much people have to translate their intentions into prompts, clicks, and keystrokes. A computer that understands enough of someone’s ongoing work could shorten the distance between deciding what to do and carrying it out while leaving the final decision with the person.

“The best suggestion should feel like it belongs exactly where you are already working,” he said. “You see the next action, decide whether it is right, and keep moving.”