Train Smarter.
Deploy Faster.
AxoLexis is a professional desktop platform for training, evaluating, and deploying machine learning models — powered by the cutting-edge SHADA algorithm, an end-to-end self-supervised hierarchical deep learning framework.
Everything You Need to Train ML Models
A complete, end-to-end platform covering every stage of the ML lifecycle — from raw data to deployed, production-ready models.
Dataset Management
Load, preview, and configure datasets with built-in preprocessing and augmentation pipelines. Supports CSV, JSON, image folders, and HuggingFace hub.
Model Configuration
Choose from Nano to XL model tiers (~150M to ~7B params). Configure architecture, hyperparameters, LoRA adapters, and optimization settings visually.
SHADA Training Pipeline
4-phase unified training: Self-supervised pre-training, Multi-task fine-tuning, Supervised fine-tuning, and Deployment optimization — all from one interface.
Real-time Evaluation
Live metrics dashboard with accuracy, loss curves, F1 score, and per-class analysis. Compare multiple runs side-by-side with beautiful charts.
Desktop Application
Native Python/Qt desktop app with a modern glassmorphism UI. No browser required — full GPU access, local data privacy, and offline operation.
Export & Deploy
Export models to ONNX, TorchScript, or HuggingFace format. Built-in quantization (INT4/INT8), pruning, and deployment optimization with one click.
RL Alignment
Optional reinforcement learning alignment with PPO (RLHF) and DPO. Fine-tune model behavior from human preference data with full KL divergence control.
Multi-modal Support
Unified encoder for both NLP and Computer Vision tasks. Train on text classification, sentiment analysis, image classification, and more from one platform.
Checkpoint Management
Initialize from any foundation model — CLIP, LLaMA, ViT, DINOv2. Full checkpoint saving, resuming, and experiment versioning built in.
From Raw Data to Deployed Model
AxoLexis guides you through every step with an intelligent wizard that adapts to your workflow and project requirements.
Upload Your Dataset
Load datasets from local files, HuggingFace Hub, or connect to cloud storage. AxoLexis auto-detects format (CSV, JSON, image folders) and provides interactive preview and statistics.
Configure Your Model
Select model tier (Nano to XL), configure architecture, set hyperparameters, and enable LoRA adapters — all from an intuitive step-by-step wizard interface.
Train with SHADA
Launch the 4-phase SHADA training pipeline. Monitor real-time loss curves, gradient norms, and learning rate schedules with live charts and progress tracking.
Evaluate & Export
Run comprehensive evaluation with accuracy, F1, confusion matrices, and benchmark comparisons. Export to ONNX, TorchScript, or HuggingFace format in one click.
Supported ML Workflows
AxoLexis covers the full machine learning lifecycle with dedicated, purpose-built workflow panels for every stage.
Multi-Phase Training
Self-supervised to Multi-task to Supervised to Deployment
AxoLexis orchestrates the full SHADA 4-phase training pipeline automatically. Start from unlabeled data with self-supervised pre-training, progress through multi-task and supervised fine-tuning, and end with deployment optimization — all configurable via the UI.
Technical Overview
AxoLexis is built on a research-grade foundation, combining the state-of-the-art SHADA algorithm with a professional desktop application interface.
Modality-Agnostic Architecture
The SHADA encoder processes both image (ConvNeXt stem + hierarchical stages) and text (transformer stages) inputs through the same unified backbone, enabling true multi-modal training from a single codebase.
Scalable from Edge to Cloud
Model tiers range from Nano (~150M params, suited for edge deployment) to XL (~7B params + MoE, requiring 32× H100 infrastructure). Train the same architecture at any scale without code changes.
Privacy-First Desktop Design
AxoLexis runs entirely locally — your data never leaves your machine. Full GPU utilization, offline operation, and direct filesystem access make it ideal for sensitive research and enterprise environments.
Technology Stack
SHADA Integration
AxoLexis is built around the SHADA (Self-supervised Hierarchical Adaptive Deep Algorithm) — a unified 4-phase training framework.
Omar Alghafri
The visionary behind the SHADA algorithm and the AxoLexis application. Dedicated to pushing the boundaries of self-supervised learning and intuitive machine learning software.