Muntin

Muntin's user is two weeks into a payroll implementation, beside a consultant mapping pay groups. The subsidiary incorporated last year sits outside the quoted plan, so it stays on a spreadsheet and somebody reconciles two payroll truths by hand each month. Muntin Systems GmbH has built for that person in Leipzig since June 2021.

Muntin sits above whatever runs each entity's payroll. Its register holds one row per legal entity: pay calendar, currency, cost center tree. Each period's file is mapped onto one fixed row contract, the original kept beside it so a journal traces back to its file.

Every period Muntin runs one set of checks across every entity in a group. They catch one person on two registers, a leaver still paid, a file that never arrived. Failures queue under a named owner; the period locks only when the queue is empty.

Muntin needs machine learning in two places: reading a payroll file nobody has seen before, and matching one person across entities. A header dictionary fails on a new bureau's file, so a fine-tuned classifier reads each header with a fingerprint of its column's values. With no shared payroll ID, the duplicate check rests on an embedding model over name, date of birth and role.

A person confirms every mapping and identity match in Muntin; the models decide nothing. Below a fixed threshold no match is guessed: the row stays unresolved and the close waits. Each confirmed mapping is a labeled pair in a corpus of headers and value statistics we own and retrain the classifier on. Identity pairs are employee personal data, partitioned per customer and never pooled.

A new group's backfill, month-end and a new embedding model each bring Muntin a burst of compute. A backfill embeds and matches three closed periods from every entity in one pass, nearly every customer reconciles in the same 72 hours at month-end, and a new model means re-embedding everything. Sizing for the peak would idle capacity three weeks in four.

Most of Muntin's infrastructure bill is cloud capacity in Frankfurt, and the larger share is GPU hours. The classifier and the matcher share cloud GPU instances raised on demand for a backfill or month-end, and object storage there holds every source file beside its rows. The bill is not optimized yet: pricing is flat per group, so one arriving with forty entities and eleven bureaus loses us money.

From Q3 2027 Muntin's classifier and identity-matching index run on reserved cloud GPU capacity in Frankfurt, with on-demand hours for peaks. Azure Machine Learning retrains the column classifier in Germany West Central, Microsoft's Frankfurt region, on headers and value fingerprints only. As a German GmbH holding salaries under GDPR, we send no payroll data to a hosted model API and process none outside Frankfurt.