Research & open source

Ideas in motion.
Code that goes further.

We explore artificial intelligence, infrastructure and mathematical computation. Discover the research projects and open-source tools that grow from this work.

github.com/chronokairo

01 / Projects

From research to implementation.

Infrastructure & DevOps

NIMBUS

Research & development

Research and development of a unified dashboard for infrastructure, deployments and quota management across cloud providers.

  • Servers, networks, SSH, Docker and Kubernetes
  • Monitoring, CI/CD, deployment and backup
  • Quota management across cloud providers
  • Infrastructure core written in Rust

Coding agents

KAIDE

Research & development

Chronokairo’s coding agent with a Rust core and terminal, desktop and web interfaces. It explores local models, RAG and agent coordination.

  • Rust core and Tauri desktop application
  • Terminal and web experiences
  • RAG with Ollama and Supabase
  • Control plane for virtual agents

Mathematical computation & verification

AXIOM

View repository

Mathematical infrastructure for agents, researchers, students and developers. A unified MCP interface connects specialized engines, with traceable results and independent verification.

  • Symbolic algebra and numerical computation
  • Integrations with SymPy, SageMath, Julia and Lean 4
  • Distinguishes proofs, approximations, refutations and inconclusive results
  • Open source under the MIT license

Inference runtime

RELIC

View repository

A minimalist GGUF runtime for OpenCL 1.2 and 3.0, focused on local models running on legacy GPUs, integrated GPUs and hybrid systems.

  • Memory budgets of 1–4 GB VRAM
  • FP16, Q8_0 and Q4_0
  • Statically planned buffers
  • Inference research on constrained hardware

02 / Research directions

What we are investigating.

01

FlashAttention on constrained hardware

Tiling, fused softmax and online normalization adapted to OpenCL and memory-constrained pipelines.

02

Distributed speculative decoding

An integrated GPU drafts tokens with a small model while a dedicated GPU verifies batches.

03

Extreme quantization

Sub-4-bit methods such as NF4, AWQ, INT2 and INT3 adapted to older GPUs’ memory bandwidth.

04

Mixture-of-Experts with limited VRAM

Dynamic expert preloading to run MoE models on a single workstation GPU.