We build tools that make mathematical knowledge robust — verifiable by machines, stress-tested against adversarial inputs, and structured for retrieval.
Dimension Reducers exists to make mathematical knowledge robust — verifiable by machines, stress-tested against adversarial inputs, and structured for retrieval. We build the tools that referee papers, torture-test LLMs, and turn mathematical corpora into clean, verified training data.
The math is not decoration. It is the product.
We are not a general-purpose AI company. We are a mathematics robustification startup with deep roots in research and a track record of deploying verified tools at scale.
Our methods are grounded in peer-reviewed mathematics — not heuristics. When we say a system works, we mean it in the sense a mathematician means it: we can show why.
We have shipped production systems — semantic search over 700K research papers, real-time sentiment platforms, automated proof auditing — not prototypes that stall in review.
We benchmark the latest LLMs, track formalization tools like Lean4, and maintain active research programs. We do not advise on technology we have not already tested.
Four areas where deep mathematical expertise becomes deployed, verified tooling.
Our flagship DiRe-JAX library extends UMAP with performance and accuracy improvements grounded in differential geometry. We also build custom graph embedding pipelines (GraphEm) that preserve topological structure for knowledge graphs, biological networks, and social network analysis.
Semantic search that exploits mathematical structure — theorems, definitions, proof dependencies — not just text similarity. Our engine indexes 700,000+ arXiv papers and returns AI-synthesized answers with citations. The architecture is designed to generalize across mathematical corpora.
Automated auditing of mathematical claims in papers and LLM outputs. Adversarial stress-testing that generates hard edge cases, finds counterexamples, and produces verified problem–solution pairs for LLM training. Includes benchmarking of frontier models on formal proof generation (Lean4).
The torture testing and refereeing pipelines produce a byproduct that may be more valuable than the audits themselves: verified, adversarially-generated mathematical problem–solution pairs. Clean training data for LLMs, grounded in proofs — not scraped from the internet and hoped for the best.
Dimension Reducers was founded by Igor Rivin, a mathematician and technologist whose career has moved between the world's leading academic institutions and the commercial frontier of computation. That combination — theoretical depth plus shipping culture — is the firm's central value proposition.
As Professor of Mathematics at Temple University and former Regius Professor at the University of St. Andrews, Igor's academic work has generated over 3,300 citations, spanning hyperbolic geometry, spectral graph theory, algebraic number theory, and computational group theory. He has held positions at the Institute for Advanced Study, Caltech, and the Institut des Hautes Études Scientifiques.
In industry, he served as Director of Advanced Development at Wolfram Research, where he built core components of the Mathematica kernel — including its compute and graphics engines. As Chief Research Officer at Cryptos Fund, he co-created the CCi30 Cryptocurrencies Index and led the application of mathematical methods to quantitative finance.
Today, through Dimension Reducers, he leads a portfolio of projects that combine published research with deployed software — from mathematics torture testing and automated refereeing to structured RAG over mathematical literature. See the full project portfolio.
We work with AI labs, publishers, and research groups who need mathematical knowledge they can trust — for training, for refereeing, for retrieval. If that's you, we'd like to hear about it.
Start the conversation