Any Docker workload

HyperOptimizer is not tied to one framework. If your program can run in Docker and expose parameters through CLI flags, it can be optimized.

The universal contract

1. Receive: Parse arguments such as --hpo-threshold=0.7.
2. Run: Execute one model training job, simulation, backtest, or data pipeline run.
3. Measure: Compute the objective and supporting metrics.
4. Print: Emit hpo.metrics.* lines to stdout.

Good workload fits

Machine learning

Tune learning rates, regularization, model dimensions, and data settings.

Trading research

Tune strategy thresholds, position sizing, and risk controls.

Simulations

Tune model assumptions, scenario settings, and objective tradeoffs.

Pipelines

Tune thresholds, batch sizes, and scoring rules for repeatable jobs.

LLM workflows

Tune prompts, retrieval settings, reranking weights, and evaluation thresholds.

Custom algorithms

Tune any executable program that reports a numeric result.

Integration checklist

• The workload can run headlessly in a container.
• Every tunable value can be passed through a --hpo-* flag (or plain --* if you configure that mode).
• The program emits at least one metric for every successful trial.
• Invalid parameter combinations fail clearly or emit a penalty metric.
• The image is published to your org Images registry before you create an experiment.

Start with the Quickstart.