Recursive self-improvement
We’re researching how Model Factory can improve itself across models, research agents, training, inference and GPU execution.
Model Factory
Work with our forward deployed researchers, research agents and GPU infrastructure to develop your model, run experiments and serve it in production.
Forward deployed researchers
Research agents
GPU infrastructure
Research, training
and inference.
Built for autonomous research and recursive self-improvement.
RESEARCH BRIEF
Research objective
Target behavior
Data
Training environment
Define the problem
Our forward deployed researchers define the model’s target behavior with your team, prepare data and develop or source training environments where needed.
SFT setup · Python
trainer = SFTTrainer( model=model, train_dataset=train_data, eval_dataset=validation_data, )
Train and evaluate
Researchers and agents run supervised fine-tuning or reinforcement learning experiments on GPU infrastructure, then evaluate results against the behavior your application needs.
Optimized model
Hedra
/
Your infrastructure
Optimize and serve
We optimize the model and inference stack for your hardware and workload, then work with your team to deploy on Hedra or your infrastructure.
Learn from production
We use production results to see where the model falls short and decide what to test next.
Test changes across models, research agents, training, inference and GPU execution, then carry what works into the next cycle.
Advancing models
and infrastructure.
Mechanistic interpretability
We investigate how models represent concepts and how to steer them without disrupting other behavior.
Model training
Supervised fine-tuning and reinforcement learning to develop the behavior your application needs.
AI researchers
Our researchers set the goals and decide which changes to keep. Research agents run experiments and evaluate the results.
