Kepler

Kepler is an agentic scientific research platform designed to carry a research question through the full workflow, from literature discovery and experimental design to analysis, writing, and review. Scientists can stay hands-on or delegate stages to agents that discover tools and data, run quantitative analyses, draft manuscripts, and perform adversarial review. Automated tool discovery and integration with Quasar or internal knowledge bases support reproducible, publication-ready research.

Problem

A model can write a confident result that a careful study would not stand behind. Leakage, an uncalibrated false-discovery rate, and a manuscript drafted before the statistics are finished all look the same on the page.

Approach

Kepler is a desktop research OS. The shell is Electron, React, and Fastify: a canvas of live tiles and real terminals, with Monaco for code, and a builder that can arrange that canvas and launch those terminals. It drives a bundled Python co-scientist engine.

A study is a scenario. Before a run, the engine validates it against a rigor floor: a permutation null, a permutation-calibrated empirical false-discovery rate, group-aware cross-validation, and a firewall against scaffold or site leakage. Findings move up a discovery ladder — screen candidate, provisional, confirmed, claim — and a manuscript is written only for what that ladder allows. A screen candidate is never reported as confirmed. A local run needs no cloud account. Heavy GPU work is the optional path off the machine.

Outcome

Every co-scientist the app builds inherits the same gates. A result that has not cleared them cannot be stated as confirmed.