No H100? I built a crowdfunded execution layer for autoresearch

[nblintao]

autoresearch is one of the most exciting things I've seen. But there's a catch: it assumes you have that GPU. Many independent researchers don't.

autoresearch solves the iteration loop beautifully. But it runs on your machine. If you don't have an H100 sitting around, you're back to square one: renting cloud GPUs, wiring up environments, and figuring out who's paying for all those runs.

So I built ML Patron, a platform that connects ideas to execution.

How It Works

Researchers submit ML experiments. Sponsors fund the ones they find interesting. The platform handles the rest: cloud GPUs, reproducible environments, metric tracking via MLflow.

Three things it tries to solve:

• The funding gap. Most experiments don't need $10K. They need $5–50 for that first baseline. ML Patron lets anyone back an experiment with a few dollars. No committees, no proposals. If someone thinks it's worth a shot, it gets run.

• The infrastructure tax. You submit a repo, pick a GPU, set parameters. The platform handles scheduling, execution, and logging. A "dry run" verifies the pipeline first at minimal cost. No K8s YAMLs, no cluster management.

• Research continuity. Every project and run has persistent research notes and discussions. The goal is that the reasoning behind each decision, not just the metrics, stays with the experiment, not in someone's head or a compacted context window.

Agent-Native by Design

This is the part most relevant to this community: ML Patron treats AI agents as first-class users.

Everything is API-first. There's a public skill.md that describes the full platform workflow: endpoints, state machines, examples. An agent reads this one file and can immediately create projects, submit runs, fund experiments, check metrics, and write research notes. No browser, no human clicking buttons.

As a proof of concept, I had Claude Code autonomously run a baseline search on nanochat. The agent read skill.md, estimated costs, submitted runs, handled a spot preemption, and iterated on configurations across A100 and H100. Two days, ~$42, two verified baselines. I stayed on the sidelines as an advisor. The full write-up is here if you're curious.

Early Stage, Looking for Feedback

ML Patron is still young, more of a working prototype than a polished product. I don't know which parts will generalize beyond my own workflow. But I think autoresearch proved that agents can do real research when given the right loop. ML Patron is my attempt to make that loop accessible to people who don't have their own hardware.

If you're interested, come try it out at mlpatron.com. I'm happy to sponsor a few runs while I figure out the rough edges.