AI-native customer operations.
Kroonen AI builds the Libre ecosystem: AI support, local LLM interfaces, Twilio-powered phone systems, and model research for small teams that refuse bloated SaaS stacks.
From model research to real products
The same lab that trained Genesis also ships practical tools: chat, docs, phone, and customer workflows.
What We Do
AI products, communications systems, and local-first infrastructure — from prototype to production.
Pre-Training from Scratch
Training language models from scratch on consumer GPUs. Custom architecture, tokenizer, and 60B token multilingual dataset - no datacenter required.
Model Fine-Tuning
Custom fine-tuning of open-weight models for your specific use case, brand voice, and domain expertise.
Safety Evaluation
Comprehensive model evaluation aligned with frontier lab methodologies. Red team testing and benchmark analysis.
Applications & Communications
Custom web, mobile, and telephony applications with AI integration. React, TypeScript, Twilio, Cloudflare Workers, and native iOS.
Red Team & Security
Adversarial testing of AI systems and infrastructure. Prompt injection, jailbreak resistance, model extraction, and CBRN safety assessments.
AI Agents & Orchestration
Design and deployment of AI agent systems that work alongside your team. Multi-model orchestration, local deployment, and custom integrations.
Featured Projects
The Libre ecosystem: support chat, local AI, and programmable communications.
Latest from the Blog
Technical deep-dives from the Genesis training run.
Genesis 1B: Run 2 Extended - 80,000 Steps
Run 2 extended to 80,000 steps (~42B tokens). Step 60,000/80,000. ETA April 24, 2026.
Genesis 1B, Run 2: 3x Throughput, Same Hardware
Redesigning Genesis 1B from 20 to 32 layers. Same param count, same GPUs, 3x training throughput.
The Optimizer State Bug: A Silent Failure in DCP Resume
A silent AdamW state bug during Run 1 that produced a false recovery on poisoned weights.
The Genesis Manifesto
Data sovereignty, constitutional alignment, and why the future of AI is local, private, and personality-first.
Fixing FSDP Checkpoint Deadlocks on 2x RTX 4090
How DCP sharded checkpoints and CPU-offload resume fixed deadlocks on consumer GPUs without NVLink.
Mapping the Mind of Qwen 3.5 9B
A sparse autoencoder for mechanistic interpretability: zero dead features, 16,384 dimensions.
Let's build together
Tell us about your project and we'll get back to you within 24 hours.
Or reach out directly