An autonomous research engine for one person's life. It collects passively, finds patterns, designs pre-registered studies — and runs each analysis once, by rules locked before the result exists. Then it remembers what it learned.
A real run, on the founder's real data. One exposed night cannot decide anything — and the engine says so instead of guessing. Even the interval's own calibration (~87% real coverage, measured on simulated data) ships with every report.
Scan, hypothesize, pre-register, test, interpret, remember — the cycle a good analyst runs once, the engine runs every morning. The only decision left to a human is locking the study. That one stays human on purpose.
The scanner finds deviations, clusters and trends, and proposes study questions — without revealing effect sizes. Measuring the effect is the study's job; seeing it early would corrupt the test.
Insights apps correlate: "your glucose rose after pasta." Selvra pre-registers a hypothesis, locks the decision rule, runs the analysis once, and reports a confidence interval you can hold it to.
Every study leaves a lesson: question, verdict, next step. The engine reads its memory before designing, so each study builds on what is known instead of repeating it. Value compounds with time.
Glucose every five minutes, sleep and steps from a wearable, events from a calendar. Exposures can be derived from the data itself — and the design review checks each one against confounders like day type before it may be used. Manual logging shrinks to seconds a day.
The scanner proposes questions; the design layer drafts a full pre-registration — thesis, falsification, negative control, smallest effect that matters. A design review gates it. You lock it. That is the whole job.
When the data suffices, the single confirmatory analysis runs by the locked rule — automatically. The verdict is interpreted in plain language, prioritised against open questions, and remembered for the next design.
Give the engine a passive data source, a measurable goal and an event log, and the same loop runs. Health first — where wrong conclusions cost the most, so the rigor is earned where it is hardest.
The research engine on real CGM data: Dexcom readings every five minutes, wearable sleep and steps, pre-registered studies with locked decision rules. "Does training lower my overnight glucose?" gets a confidence interval, not a guess — validated value-for-value against reference implementations.
Identifies what makes a musician's voice distinct. Separates signature from creative trap. Designed for the protocol from inception — Selvra integration as foundation, not feature. Awaits Python SDK port.
The same loop on company data: ERP transactions as the stream, margin and churn as outcomes, price changes and campaigns as exposures. Pre-registered answers to "did it actually work?" — after the engine has earned its rigor in health.
Pre-registration, locking, a single confirmatory run per study — the discipline of clinical research, solved once.
Block bootstrap for autocorrelated days, randomisation tests for experiments, power estimated from your own data.
Every draft is checked against positivity, carryover and power on your own data before it can be locked.
Every claim traceable. Every action logged. Every result reproducible. Trust is structural, not promised.
150+ automated checks on every push: data contracts, an ontology that forces every metric to state what it does not measure, reproducibility.
Fetch, scan, run what is ready, interpret, prioritise — scheduled every morning. The analyst that never sleeps.
Selvra is in private alpha, running daily on real data. Whichever side you come from — reach out.