tools.halowerk.com

Groups the access records you pass into clients and sorts each in two stages: person or machine, then, for a machine, the purpose. Signals: cadence and its scatter, path shape and consecutive numeric runs, asset share, referer, method mix, not-found share, robots.txt, user agent product token. Each signal returns what was observed, its weight and why. A lead under two points stays unknown. Nothing is stored, no raw IP is returned: IPv4 cut to /24, IPv6 to /48.

Runs three detectors over one series and reports what each saw: a robust z score on median and median absolute deviation, a robust seasonal-trend decomposition with a z score on the residual, and an isolation forest over value and step-to-previous. Every score is returned per point, so callers set their own threshold instead of trusting a verdict. The forest draws from a fixed seed: same request, same result. Under three full cycles the seasonal figure is marked as not robust.

Turns SenML (JSON or CBOR) and LwM2M TLV into one record shape: resolved name, unit, absolute timestamp, typed value. Optionally puts every series on a shared time axis with a chosen interval and fill rule (last value, bucket mean, linear). Works only on the payload you pass: no device is contacted, nothing is stored, the same request returns the same bytes. TLV carries no type on the wire; unmapped resource ids come back as opaque base64 with a warning instead of a guess.

Walks the events you pass through the ordered steps you name and reports per step: how many arrived, conversion from the previous step and from the first, drop-off in users and as a rate, and time to the next step as median, p90 and mean. A step counts only if it happened after the one before and inside the window. Where events carry a segment, each is tested against the largest with the same two-proportion test the A/B endpoint uses. Nothing is stored.

Builds the cohort table from the events you pass: users grouped by the period of their first event, counted again in each following period. Returns the table per cohort, a retention curve across all cohorts, and, where events carry a value, cumulative value and value per user. The curve counts only cohorts old enough for a given period, so a young cohort is not scored as churn. Day, week and month are real calendar periods. Nothing is fetched, nothing is stored.

Reads one conversion experiment two ways at once. Frequentist: two-proportion test, p value, confidence interval on the absolute lift, and the sample size the requested minimum detectable effect needs at the requested power. Bayesian: Beta-Binomial posteriors, probability the variant is better, expected loss either way, credible intervals - closed form, not sampled, so the numbers repeat exactly. looks_taken carries the adjusted threshold and the rate you actually ran at.