Fill gaps in weather station data.

AWS outages leave holes in temperature, humidity and radiation series that break degree-day sums, evapotranspiration estimates and QA/QC reporting.

How it works

Diurnal and seasonal cycles plus the variables that kept recording carry a lot of the missing signal. Gaplad's gradient-boosted trees learn those relationships from your own record and validate every gap-length class against held-out observations before filling anything.

• Quality scan first — error codes, spikes, frozen sensors and re-installed sensors are found before anything is modelled.
• Validated before filled — synthetic gaps are carved from your observed temperature, humidity or wind record, filled blind, and scored against held-out truth. The winning method per gap length fills the real gaps.
• Honest by construction — short gaps use interpolation because it wins there; unfillable gaps stay unfilled with the reason stated; every filled value is flagged, with uncertainty bands from measured residuals.

Your data never leaves your browser

Error codes (-9999 and friends), frozen sensors and spikes are detected first, so the model never trains on garbage. The model trains in a background thread on your machine — there is no upload, no account, and nothing to get security clearance for.

Free tier runs the full pipeline on up to 20,000 time steps (about two years of hourly data) with a watermark line on exports. A one-time licence lifts the size limit and cleans the exports. Exported CSVs flag every filled value; the trust report is a single HTML file you can attach to a deliverable.