Research workspace
Define the objective, evaluation function, action space, datasets, constraints, compute budget, and stopping conditions in a versioned project.
- Dataset and split management
- Custom metrics and constraints
- Action-space configuration
One place to operate an entire autoresearch program.
Configure the search, connect data, allocate GPU compute, follow experiments as they run, and collect the trained models and reports Reverie produces.
Inside the platform
From project definition to delivered model.
Autoresearch engine
Reverie proposes candidate directions, trains models, reads evaluation results, and decides which parts of the search space to explore next.
- Iterative experiment planning
- Parallel candidate evaluation
- Budget-aware stopping rules
GPU execution layer
Launch and monitor accelerator-backed training without turning the research project into an infrastructure project.
- Isolated training environments
- GPU job orchestration
- Logs, checkpoints, and recovery
Experiment lineage
Follow how each model was produced, what it changed, how it scored, and which prior result informed the next experiment.
- Search-tree history
- Comparable metrics and artifacts
- Reproducible model configurations
Model and report registry
Review the strongest candidates and export the complete package your team needs to validate and use the result.
- Trained model artifacts
- Evaluation summaries
- Generated research reports
Controls for your team
Set boundaries before the run, inspect progress while it works, and pause or redirect a project when domain judgment is needed.
- Project budgets and permissions
- Pause and intervention controls
- Complete activity history
A run inside Reverie
See the research, not just the final score.
The platform keeps the full path visible—from the first baseline through failed branches and promising candidates to the delivered model.
- Baseline
Establish the starting metric and validate the project environment.
- Explore
Test diverse model directions across the configured action space.
- Refine
Allocate more compute to promising branches and improve them iteratively.
- Deliver
Package trained models, artifacts, comparisons, and the final report.