hiloop
The Insight
A growing set of companies wants to move off closed frontier models onto open-weight ones for cost, performance, personalization, or regulatory reasons. But picking an open-weight model isn’t enough, one needs to make it perform for a specific task. That requires post-training, and it isn't a one-time project: as user behavior shifts and new base models ship every few months, the model must be re-adapted continuously to stay ahead.
Today that work is slow and specialized, spanning data generation, experimentation, evaluation, and model optimization – the kind of thing only a dedicated research team can run. hiloop's bet is that any product team should be able to do it repeatedly, as their application, data, and users evolve.
The Product
hiloop is building a ‘Cursor for post-training’: a harness that helps application-layer AI companies iterate on open-weight models without having to build the entire post-training stack themselves.
It brings the full loop across generating training data, running experiments, evaluating results, and optimizing the model against a team's own definition of "good" into one place, so improving a model becomes a workflow rather than a project.
The Journey
All three founders have worked at the intersection of applied AI and developer infrastructure. Karan was an ML engineer at Reducto and led ML infrastructure at Dynamo AI, after studying math and CS at the University of Toronto. Thomas built platform systems at Reducto and was previously an engineer at Discord. Jad joins as Chief Scientist with a Cambridge PhD in compilers and formal verification.
At Reducto, Karan and Thomas repeatedly hit the problem hiloop now solves: adapting narrow-task models against hard evals, in production, for paying customers.
They went through Y Combinator with hiloop and are now building infrastructure for the next generation of companies building on open-weight models.
The Belief
It’s easy to forget how shocking it is that scaling is possible at all – in a universe which wasn’t made of reusable, interchangeable parts, mathematics would only be reasonably effective. Our company philosophy boils down to playing the particularly fortunate hand we’ve been dealt for all it’s worth: simple, compositional, and elegant components which, like toy blocks, can be put together in an infinite diversity of ways by human and agent as we seek to build ever higher toy towers.