Proximal Raises $15M for Frontier AI Data

Proximal has emerged from stealth with $15M to tackle one of frontier AI’s least glamorous but most important bottlenecks: high-quality post-training data.

What happened

The company is building infrastructure that turns signals such as agent traces, human workflows and model failures into evaluations that expose where AI systems are weak.

Those evaluations are then used to generate targeted training data for post-training and model improvement.

Why it matters

As frontier systems become more capable, the traditional model of paying humans to label examples becomes less useful in highly specialised domains. Models may already outperform the average annotator on many tasks, while the remaining weaknesses can be subtle and difficult to identify.

Proximal is therefore treating data generation as an engineering loop: detect failure modes, measure them systematically and create new data specifically to improve those weak areas.

The bigger picture

The AI stack is shifting from raw pre-training scale toward evaluation, post-training and data quality. As foundation models become more commoditised, infrastructure that helps companies adapt and improve them for specific tasks could become an increasingly valuable layer.