Lightning Rod

Ben Turtel spent a decade building ML systems at Google as an L5 in Applied AI, and somewhere along the way he noticed something broken in how we actually train AI models: we're still manually labeling data like it's 1999. So he built Lightning Rod—a platform that turns real-world documents and public data into verified training sets without the human labeling bottleneck. The secret sauce is something they call "Future-as-Label," which uses actual outcomes that have already happened as ground truth instead of paying people to annotate data by hand. Smart.

The platform pulls from news, SEC filings, Wikipedia, whatever messy historical data you've got sitting around, and transforms it into production-ready training datasets with provenance tracking, quality controls, scoring, filtering, and deduplication built in. They showed examples ranging from predicting whether Trump would impose Canadian tariffs (they nailed it—yes, 25% in March 2025) to analyzing medical literature for QA training to forecasting supply chain pressure indices. This isn't synthetic data or crowdsourced garbage—it's real-world outcomes used as labels, which makes your model actually learn from what happened, not what some annotator thought should happen.

Turtel's background is legit: ten plus years in ML/AI/NLP, founded Kazm (sold to Harvard), built Rivet, and now Lightning Rod Labs. He's a seed-stage company solving one of the most annoying bottlenecks in AI development—data labeling is expensive, slow, and inconsistent, and everyone's pretending manual labeling scales. It doesn't. Lightning Rod says "what if we just use the future as our label" and honestly, that might be the smartest framing I've heard in months.

Trusted by enterprise, government, and startups. Seed stage, based wherever they are, and likely hunting for enterprise customers who need domain-expert AI trained on their specific historical data.