AI Research · Žilina, Slovakia

Building the foundations of superintelligence

Tagib is an AI research company developing novel architectures, training methodologies, and theoretical frameworks that push the boundaries of machine intelligence — from efficient model distillation to the mathematics of creativity.

120B+
Parameters served
GPU
Accelerated inference
H100 / RTX 4090
Compute infrastructure
PhD
Research depth
Research

From theory to architecture

Our research spans the full stack of intelligence — from foundational theory on what makes systems truly creative, to production systems serving billion-parameter models on NVIDIA GPUs.

λ
Recursive Thinker Architecture
A novel approach to model distillation that replaces middle transformer layers with a single decoder layer using anchor-with-shift residual connections, repeated across multiple "thinking passes." Achieves dramatic parameter reduction while preserving reasoning capability.
distillation transformers Qwen2.5 KV cache
∞
Superintelligence Theory
Formal research into intelligence complexity — how many resources are needed to achieve a given level of intelligence. Explores the role of creativity, beauty, and Turing-completeness in designing models with low intelligence complexity, inspired by the efficiency of human cognition.
computability creativity intelligence complexity
∫
Tool-Integrated Reasoning
Production pipeline for mathematical problem-solving using large language models with code execution. Includes priority-based scheduling, entropy-weighted voting, and tiered query strategies — deployed on NVIDIA H100 GPUs via SGLang.
SGLang vLLM H100 mathematical reasoning
≈
Self-Supervised Learning for Time Series
BYOL-based representation learning on energy consumption data. Novel architectures combining convolutional, transformer, and LSTM components with FiLM conditioning — enabling few-shot transfer to downstream prediction tasks.
BYOL representation learning energy AI
⊕
Quantum Computing Simulation
High-performance quantum computer simulator in Rust with a complete gate set. Implements quantum algorithms including Grover's search, Shor's factoring, QFT, and error correction — bridging quantum theory with practical simulation.
Rust quantum algorithms simulation
Δ
LLM Steganography
Encryption and decryption tools that use large language models to generate ciphertext indistinguishable from normal text. A novel approach to cryptographic communication leveraging the generative capabilities of transformer models.
cryptography LLM steganography
Technology

GPU-accelerated AI development

We build on NVIDIA's GPU ecosystem to train, fine-tune, and deploy models — from local workstations to cloud H100 clusters.

  • Large Model Inference
    Serving 120B+ parameter models using SGLang and vLLM on NVIDIA H100 GPUs with FlashInfer attention kernels, PagedAttention, and custom JIT kernel pre-compilation.
  • Model Training & Distillation
    Training transformer models on multi-billion token math datasets. Distillation pipelines with mixed precision, quantization, and custom architectures on RTX 4090 and H100.
  • Self-Supervised Representation Learning
    GPU-accelerated BYOL and contrastive learning pipelines with Optuna hyperparameter optimization on large-scale time series data. PyTorch and TensorFlow/Keras.
  • Performance-Critical Systems
    Rust-based data processing, parallel computing with Rayon, CUDA-accelerated workloads, and high-performance inference pipelines optimized for NVIDIA hardware.

Infrastructure

  • Local GPU NVIDIA RTX 4090
  • Cloud GPU NVIDIA H100 80GB
  • CPU AMD Threadripper 7960X
  • Memory 256 GB DDR5
  • ML Frameworks PyTorch · TensorFlow · Burn
  • Inference SGLang · vLLM · FlashInfer
  • Languages Python · Rust · CUDA
Publications & Research Output

Advancing the field

Published work spanning AI, computability theory, and GPU-accelerated computing.

2024
Machine Intelligence and Creativity
D. Natingga · Research monograph on intelligence complexity, the role of beauty and creativity in superintelligence, Hebbian learning, and the search for general simple superintelligence models
2019
Embedding Theorem for the Automorphism Group of the α-Enumeration Degrees
D. Natingga · PhD thesis, University of Leeds · Computability Theory
2018
Data Science Algorithms in a Week
D. Natingga · Packt Publishing · Top 7 algorithms for scientific computing, data analysis, and machine learning
2012
An Experiment with Asymmetric Algorithm: CPU vs. GPU
S. R. Upadhyaya, D. Tóth · Springer DASFAA · GPU-accelerated parallel computing research
Team

Led by deep expertise

DN
Dr. Dávid Natingga
Founder & Chief Scientist
Mathematician and machine learning researcher with a PhD in Computability Theory from the University of Leeds and an MEng in Computing (Artificial Intelligence) from Imperial College London. Former Forward Deployed Engineer at Palantir Technologies. Published author (Packt Publishing) and published poet. Over a decade of experience building production ML systems, from GPU-accelerated deep learning pipelines to novel algorithm design. Current research focuses on superintelligence theory, efficient model architectures, and the mathematical foundations of creativity and intelligence.
PhD Computability Theory MEng AI, Imperial College Ex-Palantir Published Author NVIDIA GPU Developer Kaggle Competitor