Proximal: Building the Feedback Loops That Make AI Better

AI models have gotten remarkably good at learning from the internet. The next challenge is teaching them how to improve once the easy data runs out.

That is the problem Proximal is going after.

The company is building infrastructure that helps AI systems learn from real-world experience. It started in software engineering, where models can attempt complex coding tasks, see what worked, identify where they failed, and use that feedback to improve. In a little over a year, Proximal has grown to more than $200 million in annualized revenue working with frontier labs and enterprises on coding agents, and recently raised a $15 million seed round led by General Catalyst at a $300 million valuation.

The revenue is striking. But what interested us from the beginning was the much bigger idea underneath it.

The Data Problem Is Changing

For much of the modern AI era, better models have depended on more data. Gather examples. Label them. Show the model what a good answer looks like. Repeat at enormous scale.

That approach gets harder as the models themselves become more capable.

If an AI system can already perform a task as well as most people, simply hiring more people to demonstrate that task or grade its answers eventually stops working. The pool of humans capable of teaching the model gets smaller, while the problems required to push the model forward get harder.

Proximal believes this changes the nature of the data problem.

The limiting factor is increasingly not how many people you can recruit to create training examples. It is whether you can build systems that continuously uncover where a model is weak, create difficult problems that target those weaknesses, evaluate what happens, and turn the result into new training data.

In other words, the opportunity is to build better feedback loops.

That requires research, infrastructure and a lot of software.

It is why Proximal describes itself as a research lab focused on data rather than a traditional data-labeling company. The company treats the data layer with the same technical rigor that frontier labs apply to model architectures and training algorithms.

Software Engineering Was the Perfect Place to Start

Software is a particularly interesting proving ground because there are clear ways to measure progress.

An agent can make a change, run the code, test it, observe the result and try again. Over time, those interactions create rich information about what the model understands, where it struggles and what it needs to learn next.

Proximal built infrastructure around that loop.

Rather than relying exclusively on manually created examples, its systems can turn the behavior of coding agents and the artifacts produced by real engineering workflows into evaluations, identify capability gaps, and generate new training data targeted at those weaknesses.

This is a fundamentally different way to think about data. Instead of treating it as a static input collected before training begins, data becomes part of a continuous system of experimentation and improvement.

Proximal has also pushed toward increasingly difficult evaluations. Its FrontierSWE work, for example, tests models on long-horizon software engineering tasks where agents have much more time to work through a problem than they typically receive on conventional coding evaluations. The company’s current leaderboard shows even frontier models completing only a portion of those tasks.

That gap is the opportunity.

If agents are eventually going to work autonomously for days or weeks, taking responsibility for increasingly complicated engineering projects, they need a way to learn from problems that look more like the work they will actually be asked to do.

A Team Built for the Problem

The team was one of the main reasons we invested.

Calvin has been building companies for years. Before Proximal, he bootstrapped a logistics software company to $1.5 million in ARR and sold it for $9 million. He later studied business and computer science at USC before dropping out for Y Combinator. He brings an unusual combination of technical curiosity, commercial instinct and a willingness to move quickly when he sees a larger opportunity.

Navid was the second engineer at Cursor, giving him a front-row seat to the rapid emergence of AI-native software development. He studied at the University of Toronto before joining Cursor and eventually teaming up with Calvin and Justus to start Proximal.

Justus came to Proximal from Prime Intellect, where he spent a year building open-source reinforcement learning infrastructure and worked on projects including SYNTHETIC-1, its effort to generate verified synthetic reasoning data across math, coding and science. He was also previously a founder himself, co-founding Revideo through Y Combinator, and earlier worked on machine learning at Dynamo AI.

Together, the three founders bring a pretty unusual combination of company-building experience, coding-agent experience and deep technical work in reinforcement learning and synthetic data.

They have since assembled a small team of researchers and engineers with backgrounds from Cursor, Google DeepMind, Meta Superintelligence, Prime Intellect, Citadel and Jane Street. More than half of the team are former founders.

That combination matters. Proximal is tackling a research-heavy problem, but it also has to turn that research into production infrastructure that some of the most sophisticated AI organizations in the world are willing to rely on.

The team has shown an ability to do both.

Beyond Coding

Software engineering is only the beginning.

One reason mathematics has become such a productive domain for AI research is that the feedback loop is unusually good. There are enormous numbers of problems to work on, and in many cases an answer can be checked objectively. That makes it possible for a model to attempt something, learn whether it was right, and iterate without requiring a human expert at every step.

Most important real-world domains do not come packaged that neatly.

Drug discovery involves experimental data, biological systems and complicated measures of success. Semiconductor design involves enormous search spaces and constraints that span physics, architecture and manufacturing. Critical infrastructure often runs on decades of accumulated software whose behavior is difficult to understand, much less safely modify.

Proximal’s larger ambition is to build the infrastructure that makes similarly powerful feedback loops possible in domains like these.

The underlying system is the same: understand where a model is failing, generate the right next challenge, evaluate the outcome and turn that experience into something the model can learn from.

If that works, the market extends far beyond coding agents.

Why We Invested

When we first met Proximal, it was easy to see the immediate opportunity around reinforcement learning environments for software engineering.

What became more interesting was the realization that environments were only one expression of a much broader idea.

As AI systems become more capable, the world will need new ways to teach them.

Human-generated datasets will remain important, but there are limits to how far manual data creation can take increasingly capable models. The next generation of training infrastructure will need to extract signal from what models do, identify what they still cannot do, and automatically create the conditions that help them improve.

Proximal is building that infrastructure.

Coding has given the team an unusually strong starting point and a market that is already developing quickly. More importantly, it has given them a place to prove that the feedback loop works before expanding into harder domains.

If they succeed, Proximal will be much more than a provider of data for AI models.

It could become part of the learning infrastructure that allows those models to keep getting better long after the obvious training data has been exhausted.

We’re excited to be working with Calvin, Navid, Justus and the entire Proximal team as they build it.

Visit Proximal at https://www.proximal.so.