Divisions
Our five business divisions
Katayama Mathematical Sciences
The Mathematical Sciences Division works across formal methods, machine learning, and algorithmic optimization. Specific lines of work include verification, property-based testing, formal-methods approaches to accelerating parallel software construction, novel high-performance learning algorithms, large-scale entity resolution, and the optimization of core algorithmic primitives.
Key Business Areas
- Formal Verification
- NoGoals, a static site generator written in Lean 4, released under Apache-2.0. It builds bilingual Japanese and English websites and proves or checks, on every build, what it claims about them. This website is built with it.
- Property-Based Testing
- Hanrei, a property-based testing library, in development.
- Large-Scale Entity Resolution
- Finding the records that describe the same entity across data sources, at scale. The hard part is training data: real matches are rare, and the ways two records can differ are endless. We use a large language model to synthesize record pairs, active learning to steer that synthesis toward the cases the current model gets wrong, and then amplify the result into a training set for a fast matching model. LightGap, in development, is the system built this way.
- High-Performance Learning Algorithms
- Decisions a large language model can make, at machine speed. The small models we train on amplified data answer in 2 to 6 ms from 207 KB of memory, train on 100,000 rows in 170 ms, and update in place as labels arrive. Accuracy on our entity-resolution benchmarks is 97%.
- Algorithmic Optimization
- Core primitives measured in microseconds. Our feature engine for entity resolution estimates how rare a name or address is from frequency tables covering 14 countries and 12 languages, in about 21 µs per evaluation, and computes string similarity and Fellegi-Sunter weights at the same speed. Every comparison LightGap makes runs through it, so these microseconds decide how much the system can afford to compare.
Katayama Digital Entertainment
The Digital Entertainment Division builds works that shape how people think — through play. The work spans LLM-driven narrative games, taste-aware recommenders, and educational pieces.
Key Business Areas
- LLM-Driven Interactive Works
- Works we are designing in which the language model itself is the game state.
- Narrative Games
- Narrative games built around LLMs, in production.
- Taste-Aware Recommenders
- Apps in development that learn taste through conversation and recommend music and video.
- Educational Pieces
- The First Principle, a science arcade game, and conversational learning pieces, in production.
Katayama Life Sciences
The Life Sciences Division is a computational biology research organization. Its research areas include generative models and data infrastructure for clinical applications. The division is in formation and has no shipping products.
Key Business Areas
- Computational Biology
- Computational methods for analyzing data such as single-cell RNA-seq measurements.
- Generative Models
- Diffusion and other generative architectures for biological data.
- Clinical Data Infrastructure
- Methods for validating the integrity of clinical data.
Katayama World Music
The World Music Division aims to carry the world's traditional folk music forward through new arrangements and new works. Alongside music production, it develops speech-model infrastructure.
Key Business Areas
- Folk-Music Reinterpretation
- Traditional folk music from Japan, Taiwan, Tuva, Hawaii, Rwanda, and Bulgaria.
- Original Arrangements
- The World Folk Chiptune Album (working title), in production: chiptune arrangements of traditional folk music for The First Principle, the Digital Entertainment Division's science arcade game.
- Audio-Model Infrastructure
- Voxtral on MAX, speech-recognition and speech-synthesis infrastructure in development for Apple Silicon.
Katayama Capital
Katayama Capital was established in December 2025. Its investment thesis centers on agents that build and operate businesses. We aim to fund the first company to reach a $1 billion valuation with no human team.
Key Business Areas
- Agent-First Companies
- Companies built and run around agents, whose human team is small by design and intends to stay small.
- Verification Stack
- Lean 4 and formal-methods foundations.
- Novel Inference
- Mojo, MAX, and Apple-Silicon-native infrastructure.