Point-in-time safety
Every example is anchored to a cutoff. The model only sees what was knowable at that moment, so there is no forward leakage and a backtest reflects what would really have happened.
Research & benchmarks
Reproducible by default.
We validate graph AI on public benchmarks where the work can be checked, enforce temporal safety so backtests stay honest, and report synthetic-data results as exactly that.
Three commitments behind every number.
Held-out backtests
We evaluate on a temporal split of your own data, per prong, and report the held-out metric. Not a training-set score, and not a cherry-picked window.
Bring-your-own-data uplift
Uplift is measured on the same split: a local model versus the local model plus the graph engine. You see the delta the graph signal actually adds.
Benchmarks
Numbers you can reproduce.
On the public RelBench suite, our graph engine reaches a held-out ROC-AUC of 0.91 on rel-f1 / driver-top3, against a published GraphSAGE baseline of 0.755. RelBench is open, so the result can be reproduced rather than taken on faith.
- RelBench rel-f1 / driver-top3: 0.91 ROC-AUC (baseline 0.755)
- Domain-realistic synthetic tasks reported as synthetic, never as client wins
- Vendor and client claims, when shown, are cited and attributed
How we report numbers
Benchmarks are easy to dress up. Our standard is simple and fixed:
- Reproducible first. Public-benchmark figures, such as RelBench, are presented as reproducible and cited to the task they come from.
- Synthetic stays synthetic. Results on synthetic or sample data are labelled as such and never relabelled as a customer outcome.
- Attribution, always. Any third-party or client claim is shown with its source, side by side with our own measured result.
The goal is a number your data science team can defend in a review, not one that only looks good on a slide.
Bring a benchmark of your own.
Upload a labelled table and we will measure uplift on your temporal split, with the held-out numbers in front of you.
We will reach out within one business day.