Skip to main content

Privacy policy & legal posture

Last updated: 2026-07-11. Contact for anything on this page: teffen@sister.software.

Mailwoman's privacy story is architectural rather than procedural: the software is designed so that personal data never reaches us in the first place. This page statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. that plainly, covers the few surfaces we do operate, and lists the legal frameworks the design addresses.

The software collects nothing​

The Mailwoman libraries, CLI, and self-hosted API servers contain no telemetry, no analytics, no phone-home behavior, and no query logging by default. Addresses you geocode are processed on your own machine or your own server. When you run the in-browser engine, the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' and data files are downloaded once and every query is processed inside your browser — the text you type is never transmitted anywhere.

Addresses can constitute personal data when they are linked to a person. Mailwoman's deployment modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' keeps that data wherever it already lives: on the deployer's infrastructure, under the deployer's existing legal basis. We never become a processor of it.

The surfaces we operate​

  • This website (mailwoman.sister.software) is a static site. It sets no tracking cookies and runs no analytics.
  • The browser demo downloads modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' and data artifacts from our content delivery infrastructure, then runs locally. Demo queries are processed in your browser and are not sent to us.
  • The public trial endpoint (photon.sister.software) is provided for evaluationevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error.. Like any web server, it keeps short-lived operational access logs (IP address, request path, timestamp) for rate limiting and abuse prevention. These logs are not used for analytics, profiling, or any other purpose, are not shared, and are routinely discarded. No accounts exist; no cookies are set.

Because we hold no user accounts and no query history, data-subject requests (access, deletion, portability) have a short answer: there is nothing on file to return or delete beyond the transient operational logs described above. Requests and questions are welcome at the contact address.

  • GDPR (EU) / UK GDPR: the software enables compliance by data minimization — personal data stays on the deployer's infrastructure and none reaches us. For the surfaces we operate, the only processing is the operational logging described above (legitimate interest, short retention). We are not a processor of any deployer's data.
  • CCPA/CPRA (California) and similar regimes: we collect no personal information and sell none; there is nothing to opt out of.
  • ePrivacy (cookies): no tracking cookies on any surface we operate.
  • EU AI Act: Mailwoman's modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' is a small, non-generative tokentokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words. classifier used for address parsingaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. — a limited-scope, transparent system. We publish model cardsmodel cardA JSON metadata file (model-card.json) shipped with each weights bundle. It declares the model version, lineage, label set, required inference channels (anchor, gazetteer), calibration data, and training provenance., trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input.-data lineage, and evaluationevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. methodology for every released modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' (methodology, data provenance), which aligns with the Act's transparency expectations for AI systemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed..
  • Data-source licensing: every dataset we redistribute is used within its license, with per-source terms documented on the data licensing & provenance page — including the deliberate exclusion of share-alike sources from published artifacts pending legal review.
  • Accessibility: the documentation site and demo aim for WCAG 2.1 AA; no formal certification has been performed yet, and an independent accessibility review is part of our roadmap. We stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. this as an aspiration rather than a claimed compliance.

What we will never add​

No advertising, no sale of data, no query harvesting for modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input.. If any surface we operate ever changes what it collects, this page changes first, with a dated note.