loading...
Autoscience's AI agent Carl published a peer-reviewed paper and placed in a Kaggle competition against 3,300 teams — now it's building ML models for Fortune 500s autonomously.
ENTRY ANGLES
Build evaluation infrastructure for autonomous ML research agents (reproducible experiment environments, quality gates, regression detection) · Create vertical-specific deployment packages (fraud detection, demand forecasting, churn prediction) on top of Autoscience's autonomous ML platform
VERTICALS
CAPABILITIES
ML engineering, Experiment infrastructure design, Vertical domain expertise
Autoscience built Carl, an AI agent that does machine learning research autonomously. Carl reads papers, forms hypotheses, writes code, runs experiments, and submits results. No human in the loop.
In December 2025, Carl entered the Kaggle Santa 2025 competition. It placed silver — top tier among 3,300 competing teams. In January 2026, Carl's paper was accepted at the ICLR 2025 workshop, credited as the sole non-human author. The paper documented Carl's own research methodology.
The company raised a $14M seed round in March 2026, led by General Catalyst, with Toyota Ventures, Perplexity Fund, S32, and MaC Ventures participating. Their commercial offer: Fortune 500 companies hire Autoscience to train specialized ML models without assembling data science teams.
The bottleneck in enterprise ML isn't compute — it's ML researchers. A Fortune 500 company that wants a custom fraud detection model needs to hire data scientists, set up pipelines, run experiments, iterate, validate. That process takes months and costs seven figures.
Carl compresses that timeline. If it can autonomously produce research competitive with human teams — and the Kaggle silver and ICLR acceptance are credible evidence — then the cost curve for specialized model development shifts dramatically.
The $14M seed from General Catalyst is notable: GC doesn't write seed checks without conviction about category size. The investor syndicate (Toyota Ventures, Perplexity Fund) signals this isn't just an AI research novelty — there's industrial and infrastructure demand behind it.
Two practical entry angles.
**Evaluation infrastructure.** Carl needs benchmarks to know if its experiments are improving. The tooling for autonomous agent evaluation — reproducible ML experiment environments, automated quality gates, regression detection — barely exists. Builders who create reliable evaluation harnesses for autonomous research agents will find early customers in companies like Autoscience and its peers.
**Vertical deployment partnerships.** Autoscience targets Fortune 500 generically. But the repeatable ML problems — fraud detection at banks, demand forecasting at retailers, churn prediction at insurers — are industry-specific. Founders who specialize Carl's approach to a single vertical and build the data pipelines, compliance wrappers, and model monitoring for that vertical can move faster than a generalist platform.