# Kishan Kumar

> Kishan Kumar, founder of KyroDB, builds AI systems for scientific discovery. Explore his work in causal learning, distributed systems, and mathematics.

Canonical profile: https://heykishan.com/
Last updated: 2026-10-14

## About

Kishan Kumar — Founder, KyroDB.

I work on building AI systems that could do discovery and accelerate science; more specifically, working on generalisation, sample efficiency & mechanistic transfer. Interests include studying time, human history/evolution and society in the age of superintelligence. You will find some of my work on distributed systems & computational mathematics too.

## Research & papers

### New Hadamard Matrix Constructions

2026 · Mathematics

60+ explicit Hadamard and skew-Hadamard matrices for previously unresolved orders, including 2092, 2148, 2284 and 2292, with exact verification certificates.

- [Code & certificates](https://github.com/vatskishan03/Hadamard)

### Root Intervention Flattening

2026 · Paper

Counterexamples and corrected conditions for quotient-graph identifiability under paired interventions in causal representation learning.

- [Paper](https://heykishan.com/papers/root-intervention-flattening.pdf)

### Monochromatic Quantum Graphs

2026 · Formal proof

Formal and computational work on the Krenn–Gu monochromatic perfect-matching problem, including an audited Lean proof of the exact complex-weight no-solution result for the six-vertex, four-color slice.

- [Proof & audit](https://github.com/vatskishan03/monochromatic-quantum-graphs)

### Hybrid Semantic Cache: Cache Invalidation for High-Dimensional Similarity Search

2026 · Research note

A correctness-first approach to similarity caching, using geometric bounds and explicit freshness semantics to avoid unnecessary invalidation.

- [Read note](https://x.com/kishanvats03/status/2037390691677184267?s=20)

### KyroBench: Evaluating Context Correctness Before AI Agents Act

2026 · Paper & benchmark

A benchmark for whether an agent’s context remains fresh, correctly scoped, resistant to pollution, and independently verifiable as source knowledge changes.

- [Paper](https://kyrobench.kyrodb.com/paper)
- [Benchmark](https://kyrobench.kyrodb.com/)
- [Code](https://github.com/KyroDB/KyroBench)

## Previous work

2024—25 · Software Engineer, Scale AI

Worked on RLHF projects, AI-agent training and evaluation, automation workflows, and coding-task model assessment.

## Profiles & contact

- [GitHub](https://github.com/vatskishan03)
- [LinkedIn](https://www.linkedin.com/in/kishan-kumar03/)
- [X](https://x.com/kishanvats03)
- [Email](mailto:kishan@kyrodb.com)
- [Company](https://www.kyrodb.com/)